diff --git a/Ryujinx.Ava/AppHost.cs b/Ryujinx.Ava/AppHost.cs index 242c84e7..eb22b39e 100644 --- a/Ryujinx.Ava/AppHost.cs +++ b/Ryujinx.Ava/AppHost.cs @@ -171,6 +171,11 @@ namespace Ryujinx.Ava ConfigurationState.Instance.Graphics.AspectRatio.Event += UpdateAspectRatioState; ConfigurationState.Instance.System.EnableDockedMode.Event += UpdateDockedModeState; ConfigurationState.Instance.System.AudioVolume.Event += UpdateAudioVolumeState; + ConfigurationState.Instance.System.EnableDockedMode.Event += UpdateDockedModeState; + ConfigurationState.Instance.System.AudioVolume.Event += UpdateAudioVolumeState; + ConfigurationState.Instance.Graphics.AntiAliasing.Event += UpdateAntiAliasing; + ConfigurationState.Instance.Graphics.ScalingFilter.Event += UpdateScalingFilter; + ConfigurationState.Instance.Graphics.ScalingFilterLevel.Event += UpdateScalingFilterLevel; _gpuCancellationTokenSource = new CancellationTokenSource(); } @@ -193,6 +198,17 @@ namespace Ryujinx.Ava } } } + private void UpdateScalingFilterLevel(object sender, ReactiveEventArgs e) + { + _renderer.Window?.SetScalingFilter((Graphics.GAL.ScalingFilter)ConfigurationState.Instance.Graphics.ScalingFilter.Value); + _renderer.Window?.SetScalingFilterLevel(ConfigurationState.Instance.Graphics.ScalingFilterLevel.Value); + } + + private void UpdateScalingFilter(object sender, ReactiveEventArgs e) + { + _renderer.Window?.SetScalingFilter((Graphics.GAL.ScalingFilter)ConfigurationState.Instance.Graphics.ScalingFilter.Value); + _renderer.Window?.SetScalingFilterLevel(ConfigurationState.Instance.Graphics.ScalingFilterLevel.Value); + } private void ShowCursor() { @@ -345,6 +361,11 @@ namespace Ryujinx.Ava } } + private void UpdateAntiAliasing(object sender, ReactiveEventArgs e) + { + _renderer?.Window?.SetAntiAliasing((Graphics.GAL.AntiAliasing)e.NewValue); + } + private void UpdateDockedModeState(object sender, ReactiveEventArgs e) { Device?.System.ChangeDockedModeState(e.NewValue); @@ -411,6 +432,9 @@ namespace Ryujinx.Ava ConfigurationState.Instance.Graphics.AspectRatio.Event -= UpdateAspectRatioState; ConfigurationState.Instance.System.EnableDockedMode.Event -= UpdateDockedModeState; ConfigurationState.Instance.System.AudioVolume.Event -= UpdateAudioVolumeState; + ConfigurationState.Instance.Graphics.ScalingFilter.Event -= UpdateScalingFilter; + ConfigurationState.Instance.Graphics.ScalingFilterLevel.Event -= UpdateScalingFilterLevel; + ConfigurationState.Instance.Graphics.AntiAliasing.Event -= UpdateAntiAliasing; _topLevel.PointerMoved -= TopLevel_PointerMoved; @@ -788,6 +812,10 @@ namespace Ryujinx.Ava Device.Gpu.Renderer.Initialize(_glLogLevel); + _renderer?.Window?.SetAntiAliasing((Graphics.GAL.AntiAliasing)ConfigurationState.Instance.Graphics.AntiAliasing.Value); + _renderer?.Window?.SetScalingFilter((Graphics.GAL.ScalingFilter)ConfigurationState.Instance.Graphics.ScalingFilter.Value); + _renderer?.Window?.SetScalingFilterLevel(ConfigurationState.Instance.Graphics.ScalingFilterLevel.Value); + Width = (int)_rendererHost.Bounds.Width; Height = (int)_rendererHost.Bounds.Height; diff --git a/Ryujinx.Ava/Assets/Locales/en_US.json b/Ryujinx.Ava/Assets/Locales/en_US.json index b7d1e02b..db8d2424 100644 --- a/Ryujinx.Ava/Assets/Locales/en_US.json +++ b/Ryujinx.Ava/Assets/Locales/en_US.json @@ -626,6 +626,16 @@ "Recover": "Recover", "UserProfilesRecoverHeading" : "Saves were found for the following accounts", "UserProfilesRecoverEmptyList": "No profiles to recover", + "GraphicsAATooltip": "Applies anti-aliasing to the game render", + "GraphicsAALabel": "Anti-Aliasing:", + "GraphicsScalingFilterLabel": "Scaling Filter:", + "GraphicsScalingFilterTooltip": "Enables Framebuffer Scaling", + "GraphicsScalingFilterLevelLabel": "Level", + "GraphicsScalingFilterLevelTooltip": "Set Scaling Filter Level", + "SmaaLow": "SMAA Low", + "SmaaMedium": "SMAA Medium", + "SmaaHigh": "SMAA High", + "SmaaUltra": "SMAA Ultra", "UserEditorTitle" : "Edit User", "UserEditorTitleCreate" : "Create User" } diff --git a/Ryujinx.Ava/UI/ViewModels/SettingsViewModel.cs b/Ryujinx.Ava/UI/ViewModels/SettingsViewModel.cs index 36b37b0f..7045c9ed 100644 --- a/Ryujinx.Ava/UI/ViewModels/SettingsViewModel.cs +++ b/Ryujinx.Ava/UI/ViewModels/SettingsViewModel.cs @@ -45,6 +45,8 @@ namespace Ryujinx.Ava.UI.ViewModels private KeyboardHotkeys _keyboardHotkeys; private int _graphicsBackendIndex; private string _customThemePath; + private int _scalingFilter; + private int _scalingFilterLevel; public event Action CloseWindow; public event Action SaveSettingsEvent; @@ -153,6 +155,8 @@ namespace Ryujinx.Ava.UI.ViewModels public bool IsSDL2Enabled { get; set; } public bool EnableCustomTheme { get; set; } public bool IsCustomResolutionScaleActive => _resolutionScale == 4; + public bool IsScalingFilterActive => _scalingFilter == (int)Ryujinx.Common.Configuration.ScalingFilter.Fsr; + public bool IsVulkanSelected => GraphicsBackendIndex == 0; public bool UseHypervisor { get; set; } @@ -179,6 +183,18 @@ namespace Ryujinx.Ava.UI.ViewModels public int AudioBackend { get; set; } public int MaxAnisotropy { get; set; } public int AspectRatio { get; set; } + public int AntiAliasingEffect { get; set; } + public string ScalingFilterLevelText => ScalingFilterLevel.ToString("0"); + public int ScalingFilterLevel + { + get => _scalingFilterLevel; + set + { + _scalingFilterLevel = value; + OnPropertyChanged(); + OnPropertyChanged(nameof(ScalingFilterLevelText)); + } + } public int OpenglDebugLevel { get; set; } public int MemoryMode { get; set; } public int BaseStyleIndex { get; set; } @@ -192,6 +208,16 @@ namespace Ryujinx.Ava.UI.ViewModels OnPropertyChanged(nameof(IsVulkanSelected)); } } + public int ScalingFilter + { + get => _scalingFilter; + set + { + _scalingFilter = value; + OnPropertyChanged(); + OnPropertyChanged(nameof(IsScalingFilterActive)); + } + } public int PreferredGpuIndex { get; set; } @@ -365,6 +391,9 @@ namespace Ryujinx.Ava.UI.ViewModels AspectRatio = (int)config.Graphics.AspectRatio.Value; GraphicsBackendMultithreadingIndex = (int)config.Graphics.BackendThreading.Value; ShaderDumpPath = config.Graphics.ShadersDumpPath; + AntiAliasingEffect = (int)config.Graphics.AntiAliasing.Value; + ScalingFilter = (int)config.Graphics.ScalingFilter.Value; + ScalingFilterLevel = config.Graphics.ScalingFilterLevel.Value; // Audio AudioBackend = (int)config.System.AudioBackend.Value; @@ -447,6 +476,9 @@ namespace Ryujinx.Ava.UI.ViewModels config.Graphics.ResScaleCustom.Value = CustomResolutionScale; config.Graphics.MaxAnisotropy.Value = MaxAnisotropy == 0 ? -1 : MathF.Pow(2, MaxAnisotropy); config.Graphics.AspectRatio.Value = (AspectRatio)AspectRatio; + config.Graphics.AntiAliasing.Value = (AntiAliasing)AntiAliasingEffect; + config.Graphics.ScalingFilter.Value = (ScalingFilter)ScalingFilter; + config.Graphics.ScalingFilterLevel.Value = ScalingFilterLevel; if (ConfigurationState.Instance.Graphics.BackendThreading != (BackendThreading)GraphicsBackendMultithreadingIndex) { diff --git a/Ryujinx.Ava/UI/Views/Settings/SettingsGraphicsView.axaml b/Ryujinx.Ava/UI/Views/Settings/SettingsGraphicsView.axaml index fb30fb7f..8e4122f3 100644 --- a/Ryujinx.Ava/UI/Views/Settings/SettingsGraphicsView.axaml +++ b/Ryujinx.Ava/UI/Views/Settings/SettingsGraphicsView.axaml @@ -7,6 +7,7 @@ xmlns:ui="clr-namespace:FluentAvalonia.UI.Controls;assembly=FluentAvalonia" xmlns:locale="clr-namespace:Ryujinx.Ava.Common.Locale" xmlns:viewModels="clr-namespace:Ryujinx.Ava.UI.ViewModels" + Design.Width="1000" mc:Ignorable="d" x:CompileBindings="True" x:DataType="viewModels:SettingsViewModel"> @@ -111,6 +112,83 @@ Minimum="0.1" Value="{Binding CustomResolutionScale}" /> + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + _scale; + set + { + _scale = MathF.Max(0.01f, value); + } + } + + public FsrScalingFilter(OpenGLRenderer renderer, IPostProcessingEffect filter) + { + Initialize(); + + _renderer = renderer; + } + + public void Dispose() + { + if (_scalingShaderProgram != 0) + { + GL.DeleteProgram(_scalingShaderProgram); + GL.DeleteProgram(_sharpeningShaderProgram); + } + + _intermediaryTexture?.Dispose(); + } + + private void Initialize() + { + var scalingShader = EmbeddedResources.ReadAllText("Ryujinx.Graphics.OpenGL/Effects/Shaders/fsr_scaling.glsl"); + var sharpeningShader = EmbeddedResources.ReadAllText("Ryujinx.Graphics.OpenGL/Effects/Shaders/fsr_sharpening.glsl"); + var fsrA = EmbeddedResources.ReadAllText("Ryujinx.Graphics.OpenGL/Effects/Shaders/ffx_a.h"); + var fsr1 = EmbeddedResources.ReadAllText("Ryujinx.Graphics.OpenGL/Effects/Shaders/ffx_fsr1.h"); + + scalingShader = scalingShader.Replace("#include \"ffx_a.h\"", fsrA); + scalingShader = scalingShader.Replace("#include \"ffx_fsr1.h\"", fsr1); + sharpeningShader = sharpeningShader.Replace("#include \"ffx_a.h\"", fsrA); + sharpeningShader = sharpeningShader.Replace("#include \"ffx_fsr1.h\"", fsr1); + + _scalingShaderProgram = CompileProgram(scalingShader, ShaderType.ComputeShader); + _sharpeningShaderProgram = CompileProgram(sharpeningShader, ShaderType.ComputeShader); + + _inputUniform = GL.GetUniformLocation(_scalingShaderProgram, "Source"); + _outputUniform = GL.GetUniformLocation(_scalingShaderProgram, "imgOutput"); + _sharpeningUniform = GL.GetUniformLocation(_sharpeningShaderProgram, "sharpening"); + + _srcX0Uniform = GL.GetUniformLocation(_scalingShaderProgram, "srcX0"); + _srcX1Uniform = GL.GetUniformLocation(_scalingShaderProgram, "srcX1"); + _srcY0Uniform = GL.GetUniformLocation(_scalingShaderProgram, "srcY0"); + _srcY1Uniform = GL.GetUniformLocation(_scalingShaderProgram, "srcY1"); + _dstX0Uniform = GL.GetUniformLocation(_scalingShaderProgram, "dstX0"); + _dstX1Uniform = GL.GetUniformLocation(_scalingShaderProgram, "dstX1"); + _dstY0Uniform = GL.GetUniformLocation(_scalingShaderProgram, "dstY0"); + _dstY1Uniform = GL.GetUniformLocation(_scalingShaderProgram, "dstY1"); + _scaleXUniform = GL.GetUniformLocation(_scalingShaderProgram, "scaleX"); + _scaleYUniform = GL.GetUniformLocation(_scalingShaderProgram, "scaleY"); + } + + public void Run( + TextureView view, + TextureView destinationTexture, + int width, + int height, + Extents2D source, + Extents2D destination) + { + if (_intermediaryTexture == null || _intermediaryTexture.Info.Width != width || _intermediaryTexture.Info.Height != height) + { + _intermediaryTexture?.Dispose(); + var originalInfo = view.Info; + var info = new TextureCreateInfo(width, + height, + originalInfo.Depth, + originalInfo.Levels, + originalInfo.Samples, + originalInfo.BlockWidth, + originalInfo.BlockHeight, + originalInfo.BytesPerPixel, + originalInfo.Format, + originalInfo.DepthStencilMode, + originalInfo.Target, + originalInfo.SwizzleR, + originalInfo.SwizzleG, + originalInfo.SwizzleB, + originalInfo.SwizzleA); + + _intermediaryTexture = new TextureStorage(_renderer, info, view.ScaleFactor); + _intermediaryTexture.CreateDefaultView(); + } + + var textureView = _intermediaryTexture.CreateView(_intermediaryTexture.Info, 0, 0) as TextureView; + + int previousProgram = GL.GetInteger(GetPName.CurrentProgram); + int previousUnit = GL.GetInteger(GetPName.ActiveTexture); + GL.ActiveTexture(TextureUnit.Texture0); + int previousTextureBinding = GL.GetInteger(GetPName.TextureBinding2D); + + GL.BindImageTexture(0, textureView.Handle, 0, false, 0, TextureAccess.ReadWrite, SizedInternalFormat.Rgba8); + + int threadGroupWorkRegionDim = 16; + int dispatchX = (width + (threadGroupWorkRegionDim - 1)) / threadGroupWorkRegionDim; + int dispatchY = (height + (threadGroupWorkRegionDim - 1)) / threadGroupWorkRegionDim; + + // Scaling pass + float srcWidth = Math.Abs(source.X2 - source.X1); + float srcHeight = Math.Abs(source.Y2 - source.Y1); + float scaleX = srcWidth / view.Width; + float scaleY = srcHeight / view.Height; + GL.UseProgram(_scalingShaderProgram); + view.Bind(0); + GL.Uniform1(_inputUniform, 0); + GL.Uniform1(_outputUniform, 0); + GL.Uniform1(_srcX0Uniform, (float)source.X1); + GL.Uniform1(_srcX1Uniform, (float)source.X2); + GL.Uniform1(_srcY0Uniform, (float)source.Y1); + GL.Uniform1(_srcY1Uniform, (float)source.Y2); + GL.Uniform1(_dstX0Uniform, (float)destination.X1); + GL.Uniform1(_dstX1Uniform, (float)destination.X2); + GL.Uniform1(_dstY0Uniform, (float)destination.Y1); + GL.Uniform1(_dstY1Uniform, (float)destination.Y2); + GL.Uniform1(_scaleXUniform, scaleX); + GL.Uniform1(_scaleYUniform, scaleY); + GL.DispatchCompute(dispatchX, dispatchY, 1); + + GL.MemoryBarrier(MemoryBarrierFlags.ShaderImageAccessBarrierBit); + + // Sharpening Pass + GL.UseProgram(_sharpeningShaderProgram); + GL.BindImageTexture(0, destinationTexture.Handle, 0, false, 0, TextureAccess.ReadWrite, SizedInternalFormat.Rgba8); + textureView.Bind(0); + GL.Uniform1(_inputUniform, 0); + GL.Uniform1(_outputUniform, 0); + GL.Uniform1(_sharpeningUniform, 1.5f - (Level * 0.01f * 1.5f)); + GL.DispatchCompute(dispatchX, dispatchY, 1); + + GL.UseProgram(previousProgram); + GL.MemoryBarrier(MemoryBarrierFlags.ShaderImageAccessBarrierBit); + + (_renderer.Pipeline as Pipeline).RestoreImages1And2(); + + GL.ActiveTexture(TextureUnit.Texture0); + GL.BindTexture(TextureTarget.Texture2D, previousTextureBinding); + + GL.ActiveTexture((TextureUnit)previousUnit); + } + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.OpenGL/Effects/FxaaPostProcessingEffect.cs b/Ryujinx.Graphics.OpenGL/Effects/FxaaPostProcessingEffect.cs new file mode 100644 index 00000000..3a2d685b --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/FxaaPostProcessingEffect.cs @@ -0,0 +1,81 @@ +using OpenTK.Graphics.OpenGL; +using Ryujinx.Common; +using Ryujinx.Graphics.OpenGL.Image; + +namespace Ryujinx.Graphics.OpenGL.Effects +{ + internal class FxaaPostProcessingEffect : IPostProcessingEffect + { + private readonly OpenGLRenderer _renderer; + private int _resolutionUniform; + private int _inputUniform; + private int _outputUniform; + private int _shaderProgram; + private TextureStorage _textureStorage; + + public FxaaPostProcessingEffect(OpenGLRenderer renderer) + { + Initialize(); + + _renderer = renderer; + } + + public void Dispose() + { + if (_shaderProgram != 0) + { + GL.DeleteProgram(_shaderProgram); + _textureStorage?.Dispose(); + } + } + + private void Initialize() + { + _shaderProgram = ShaderHelper.CompileProgram(EmbeddedResources.ReadAllText("Ryujinx.Graphics.OpenGL/Effects/Shaders/fxaa.glsl"), ShaderType.ComputeShader); + + _resolutionUniform = GL.GetUniformLocation(_shaderProgram, "invResolution"); + _inputUniform = GL.GetUniformLocation(_shaderProgram, "inputTexture"); + _outputUniform = GL.GetUniformLocation(_shaderProgram, "imgOutput"); + } + + public TextureView Run(TextureView view, int width, int height) + { + if (_textureStorage == null || _textureStorage.Info.Width != view.Width || _textureStorage.Info.Height != view.Height) + { + _textureStorage?.Dispose(); + _textureStorage = new TextureStorage(_renderer, view.Info, view.ScaleFactor); + _textureStorage.CreateDefaultView(); + } + + var textureView = _textureStorage.CreateView(view.Info, 0, 0) as TextureView; + + int previousProgram = GL.GetInteger(GetPName.CurrentProgram); + int previousUnit = GL.GetInteger(GetPName.ActiveTexture); + GL.ActiveTexture(TextureUnit.Texture0); + int previousTextureBinding = GL.GetInteger(GetPName.TextureBinding2D); + + GL.BindImageTexture(0, textureView.Handle, 0, false, 0, TextureAccess.ReadWrite, SizedInternalFormat.Rgba8); + GL.UseProgram(_shaderProgram); + + var dispatchX = BitUtils.DivRoundUp(view.Width, IPostProcessingEffect.LocalGroupSize); + var dispatchY = BitUtils.DivRoundUp(view.Height, IPostProcessingEffect.LocalGroupSize); + + view.Bind(0); + GL.Uniform1(_inputUniform, 0); + GL.Uniform1(_outputUniform, 0); + GL.Uniform2(_resolutionUniform, (float)view.Width, (float)view.Height); + GL.DispatchCompute(dispatchX, dispatchY, 1); + GL.UseProgram(previousProgram); + GL.MemoryBarrier(MemoryBarrierFlags.ShaderImageAccessBarrierBit); + + (_renderer.Pipeline as Pipeline).RestoreImages1And2(); + + GL.ActiveTexture(TextureUnit.Texture0); + GL.BindTexture(TextureTarget.Texture2D, previousTextureBinding); + + GL.ActiveTexture((TextureUnit)previousUnit); + + return textureView; + } + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.OpenGL/Effects/IPostProcessingEffect.cs b/Ryujinx.Graphics.OpenGL/Effects/IPostProcessingEffect.cs new file mode 100644 index 00000000..7a045a02 --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/IPostProcessingEffect.cs @@ -0,0 +1,11 @@ +using Ryujinx.Graphics.OpenGL.Image; +using System; + +namespace Ryujinx.Graphics.OpenGL.Effects +{ + internal interface IPostProcessingEffect : IDisposable + { + const int LocalGroupSize = 64; + TextureView Run(TextureView view, int width, int height); + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.OpenGL/Effects/IScalingFilter.cs b/Ryujinx.Graphics.OpenGL/Effects/IScalingFilter.cs new file mode 100644 index 00000000..e1e1b2c1 --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/IScalingFilter.cs @@ -0,0 +1,18 @@ +using Ryujinx.Graphics.GAL; +using Ryujinx.Graphics.OpenGL.Image; +using System; + +namespace Ryujinx.Graphics.OpenGL.Effects +{ + internal interface IScalingFilter : IDisposable + { + float Level { get; set; } + void Run( + TextureView view, + TextureView destinationTexture, + int width, + int height, + Extents2D source, + Extents2D destination); + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.OpenGL/Effects/ShaderHelper.cs b/Ryujinx.Graphics.OpenGL/Effects/ShaderHelper.cs new file mode 100644 index 00000000..72c5a98f --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/ShaderHelper.cs @@ -0,0 +1,40 @@ +using OpenTK.Graphics.OpenGL; +using System; + +namespace Ryujinx.Graphics.OpenGL.Effects +{ + internal static class ShaderHelper + { + public static int CompileProgram(string shaderCode, ShaderType shaderType) + { + var shader = GL.CreateShader(shaderType); + GL.ShaderSource(shader, shaderCode); + GL.CompileShader(shader); + + var program = GL.CreateProgram(); + GL.AttachShader(program, shader); + GL.LinkProgram(program); + + GL.DetachShader(program, shader); + GL.DeleteShader(shader); + + return program; + } + + public static int CompileProgram(string[] shaders, ShaderType shaderType) + { + var shader = GL.CreateShader(shaderType); + GL.ShaderSource(shader, shaders.Length, shaders, (int[])null); + GL.CompileShader(shader); + + var program = GL.CreateProgram(); + GL.AttachShader(program, shader); + GL.LinkProgram(program); + + GL.DetachShader(program, shader); + GL.DeleteShader(shader); + + return program; + } + } +} diff --git a/Ryujinx.Graphics.OpenGL/Effects/Shaders/ffx_a.h b/Ryujinx.Graphics.OpenGL/Effects/Shaders/ffx_a.h new file mode 100644 index 00000000..d04bff55 --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/Shaders/ffx_a.h @@ -0,0 +1,2656 @@ +//============================================================================================================================== +// +// [A] SHADER PORTABILITY 1.20210629 +// +//============================================================================================================================== +// FidelityFX Super Resolution Sample +// +// Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved. +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files(the "Software"), to deal +// in the Software without restriction, including without limitation the rights +// to use, copy, modify, merge, publish, distribute, sublicense, and / or sell +// copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions : +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN +// THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +// MIT LICENSE +// =========== +// Copyright (c) 2014 Michal Drobot (for concepts used in "FLOAT APPROXIMATIONS"). +// ----------- +// Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation +// files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, +// modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the +// Software is furnished to do so, subject to the following conditions: +// ----------- +// The above copyright notice and this permission notice shall be included in all copies or substantial portions of the +// Software. +// ----------- +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE +// WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR +// COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, +// ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +// ABOUT +// ===== +// Common central point for high-level shading language and C portability for various shader headers. +//------------------------------------------------------------------------------------------------------------------------------ +// DEFINES +// ======= +// A_CPU ..... Include the CPU related code. +// A_GPU ..... Include the GPU related code. +// A_GLSL .... Using GLSL. +// A_HLSL .... Using HLSL. +// A_HLSL_6_2 Using HLSL 6.2 with new 'uint16_t' and related types (requires '-enable-16bit-types'). +// A_NO_16_BIT_CAST Don't use instructions that are not availabe in SPIR-V (needed for running A_HLSL_6_2 on Vulkan) +// A_GCC ..... Using a GCC compatible compiler (else assume MSVC compatible compiler by default). +// ======= +// A_BYTE .... Support 8-bit integer. +// A_HALF .... Support 16-bit integer and floating point. +// A_LONG .... Support 64-bit integer. +// A_DUBL .... Support 64-bit floating point. +// ======= +// A_WAVE .... Support wave-wide operations. +//------------------------------------------------------------------------------------------------------------------------------ +// To get #include "ffx_a.h" working in GLSL use '#extension GL_GOOGLE_include_directive:require'. +//------------------------------------------------------------------------------------------------------------------------------ +// SIMPLIFIED TYPE SYSTEM +// ====================== +// - All ints will be unsigned with exception of when signed is required. +// - Type naming simplified and shortened "A<#components>", +// - H = 16-bit float (half) +// - F = 32-bit float (float) +// - D = 64-bit float (double) +// - P = 1-bit integer (predicate, not using bool because 'B' is used for byte) +// - B = 8-bit integer (byte) +// - W = 16-bit integer (word) +// - U = 32-bit integer (unsigned) +// - L = 64-bit integer (long) +// - Using "AS<#components>" for signed when required. +//------------------------------------------------------------------------------------------------------------------------------ +// TODO +// ==== +// - Make sure 'ALerp*(a,b,m)' does 'b*m+(-a*m+a)' (2 ops). +//------------------------------------------------------------------------------------------------------------------------------ +// CHANGE LOG +// ========== +// 20200914 - Expanded wave ops and prx code. +// 20200713 - Added [ZOL] section, fixed serious bugs in sRGB and Rec.709 color conversion code, etc. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// COMMON +//============================================================================================================================== +#define A_2PI 6.28318530718 +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// CPU +// +// +//============================================================================================================================== +#ifdef A_CPU + // Supporting user defined overrides. + #ifndef A_RESTRICT + #define A_RESTRICT __restrict + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifndef A_STATIC + #define A_STATIC static + #endif +//------------------------------------------------------------------------------------------------------------------------------ + // Same types across CPU and GPU. + // Predicate uses 32-bit integer (C friendly bool). + typedef uint32_t AP1; + typedef float AF1; + typedef double AD1; + typedef uint8_t AB1; + typedef uint16_t AW1; + typedef uint32_t AU1; + typedef uint64_t AL1; + typedef int8_t ASB1; + typedef int16_t ASW1; + typedef int32_t ASU1; + typedef int64_t ASL1; +//------------------------------------------------------------------------------------------------------------------------------ + #define AD1_(a) ((AD1)(a)) + #define AF1_(a) ((AF1)(a)) + #define AL1_(a) ((AL1)(a)) + #define AU1_(a) ((AU1)(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ASL1_(a) ((ASL1)(a)) + #define ASU1_(a) ((ASU1)(a)) +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AU1 AU1_AF1(AF1 a){union{AF1 f;AU1 u;}bits;bits.f=a;return bits.u;} +//------------------------------------------------------------------------------------------------------------------------------ + #define A_TRUE 1 + #define A_FALSE 0 +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// CPU/GPU PORTING +// +//------------------------------------------------------------------------------------------------------------------------------ +// Get CPU and GPU to share all setup code, without duplicate code paths. +// This uses a lower-case prefix for special vector constructs. +// - In C restrict pointers are used. +// - In the shading language, in/inout/out arguments are used. +// This depends on the ability to access a vector value in both languages via array syntax (aka color[2]). +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR ARGUMENT/RETURN/INITIALIZATION PORTABILITY +//============================================================================================================================== + #define retAD2 AD1 *A_RESTRICT + #define retAD3 AD1 *A_RESTRICT + #define retAD4 AD1 *A_RESTRICT + #define retAF2 AF1 *A_RESTRICT + #define retAF3 AF1 *A_RESTRICT + #define retAF4 AF1 *A_RESTRICT + #define retAL2 AL1 *A_RESTRICT + #define retAL3 AL1 *A_RESTRICT + #define retAL4 AL1 *A_RESTRICT + #define retAU2 AU1 *A_RESTRICT + #define retAU3 AU1 *A_RESTRICT + #define retAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define inAD2 AD1 *A_RESTRICT + #define inAD3 AD1 *A_RESTRICT + #define inAD4 AD1 *A_RESTRICT + #define inAF2 AF1 *A_RESTRICT + #define inAF3 AF1 *A_RESTRICT + #define inAF4 AF1 *A_RESTRICT + #define inAL2 AL1 *A_RESTRICT + #define inAL3 AL1 *A_RESTRICT + #define inAL4 AL1 *A_RESTRICT + #define inAU2 AU1 *A_RESTRICT + #define inAU3 AU1 *A_RESTRICT + #define inAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define inoutAD2 AD1 *A_RESTRICT + #define inoutAD3 AD1 *A_RESTRICT + #define inoutAD4 AD1 *A_RESTRICT + #define inoutAF2 AF1 *A_RESTRICT + #define inoutAF3 AF1 *A_RESTRICT + #define inoutAF4 AF1 *A_RESTRICT + #define inoutAL2 AL1 *A_RESTRICT + #define inoutAL3 AL1 *A_RESTRICT + #define inoutAL4 AL1 *A_RESTRICT + #define inoutAU2 AU1 *A_RESTRICT + #define inoutAU3 AU1 *A_RESTRICT + #define inoutAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define outAD2 AD1 *A_RESTRICT + #define outAD3 AD1 *A_RESTRICT + #define outAD4 AD1 *A_RESTRICT + #define outAF2 AF1 *A_RESTRICT + #define outAF3 AF1 *A_RESTRICT + #define outAF4 AF1 *A_RESTRICT + #define outAL2 AL1 *A_RESTRICT + #define outAL3 AL1 *A_RESTRICT + #define outAL4 AL1 *A_RESTRICT + #define outAU2 AU1 *A_RESTRICT + #define outAU3 AU1 *A_RESTRICT + #define outAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define varAD2(x) AD1 x[2] + #define varAD3(x) AD1 x[3] + #define varAD4(x) AD1 x[4] + #define varAF2(x) AF1 x[2] + #define varAF3(x) AF1 x[3] + #define varAF4(x) AF1 x[4] + #define varAL2(x) AL1 x[2] + #define varAL3(x) AL1 x[3] + #define varAL4(x) AL1 x[4] + #define varAU2(x) AU1 x[2] + #define varAU3(x) AU1 x[3] + #define varAU4(x) AU1 x[4] +//------------------------------------------------------------------------------------------------------------------------------ + #define initAD2(x,y) {x,y} + #define initAD3(x,y,z) {x,y,z} + #define initAD4(x,y,z,w) {x,y,z,w} + #define initAF2(x,y) {x,y} + #define initAF3(x,y,z) {x,y,z} + #define initAF4(x,y,z,w) {x,y,z,w} + #define initAL2(x,y) {x,y} + #define initAL3(x,y,z) {x,y,z} + #define initAL4(x,y,z,w) {x,y,z,w} + #define initAU2(x,y) {x,y} + #define initAU3(x,y,z) {x,y,z} + #define initAU4(x,y,z,w) {x,y,z,w} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS +//------------------------------------------------------------------------------------------------------------------------------ +// TODO +// ==== +// - Replace transcendentals with manual versions. +//============================================================================================================================== + #ifdef A_GCC + A_STATIC AD1 AAbsD1(AD1 a){return __builtin_fabs(a);} + A_STATIC AF1 AAbsF1(AF1 a){return __builtin_fabsf(a);} + A_STATIC AU1 AAbsSU1(AU1 a){return AU1_(__builtin_abs(ASU1_(a)));} + A_STATIC AL1 AAbsSL1(AL1 a){return AL1_(__builtin_llabs(ASL1_(a)));} + #else + A_STATIC AD1 AAbsD1(AD1 a){return fabs(a);} + A_STATIC AF1 AAbsF1(AF1 a){return fabsf(a);} + A_STATIC AU1 AAbsSU1(AU1 a){return AU1_(abs(ASU1_(a)));} + A_STATIC AL1 AAbsSL1(AL1 a){return AL1_(labs((long)ASL1_(a)));} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ACosD1(AD1 a){return __builtin_cos(a);} + A_STATIC AF1 ACosF1(AF1 a){return __builtin_cosf(a);} + #else + A_STATIC AD1 ACosD1(AD1 a){return cos(a);} + A_STATIC AF1 ACosF1(AF1 a){return cosf(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ADotD2(inAD2 a,inAD2 b){return a[0]*b[0]+a[1]*b[1];} + A_STATIC AD1 ADotD3(inAD3 a,inAD3 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2];} + A_STATIC AD1 ADotD4(inAD4 a,inAD4 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2]+a[3]*b[3];} + A_STATIC AF1 ADotF2(inAF2 a,inAF2 b){return a[0]*b[0]+a[1]*b[1];} + A_STATIC AF1 ADotF3(inAF3 a,inAF3 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2];} + A_STATIC AF1 ADotF4(inAF4 a,inAF4 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2]+a[3]*b[3];} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 AExp2D1(AD1 a){return __builtin_exp2(a);} + A_STATIC AF1 AExp2F1(AF1 a){return __builtin_exp2f(a);} + #else + A_STATIC AD1 AExp2D1(AD1 a){return exp2(a);} + A_STATIC AF1 AExp2F1(AF1 a){return exp2f(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 AFloorD1(AD1 a){return __builtin_floor(a);} + A_STATIC AF1 AFloorF1(AF1 a){return __builtin_floorf(a);} + #else + A_STATIC AD1 AFloorD1(AD1 a){return floor(a);} + A_STATIC AF1 AFloorF1(AF1 a){return floorf(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ALerpD1(AD1 a,AD1 b,AD1 c){return b*c+(-a*c+a);} + A_STATIC AF1 ALerpF1(AF1 a,AF1 b,AF1 c){return b*c+(-a*c+a);} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ALog2D1(AD1 a){return __builtin_log2(a);} + A_STATIC AF1 ALog2F1(AF1 a){return __builtin_log2f(a);} + #else + A_STATIC AD1 ALog2D1(AD1 a){return log2(a);} + A_STATIC AF1 ALog2F1(AF1 a){return log2f(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 AMaxD1(AD1 a,AD1 b){return a>b?a:b;} + A_STATIC AF1 AMaxF1(AF1 a,AF1 b){return a>b?a:b;} + A_STATIC AL1 AMaxL1(AL1 a,AL1 b){return a>b?a:b;} + A_STATIC AU1 AMaxU1(AU1 a,AU1 b){return a>b?a:b;} +//------------------------------------------------------------------------------------------------------------------------------ + // These follow the convention that A integer types don't have signage, until they are operated on. + A_STATIC AL1 AMaxSL1(AL1 a,AL1 b){return (ASL1_(a)>ASL1_(b))?a:b;} + A_STATIC AU1 AMaxSU1(AU1 a,AU1 b){return (ASU1_(a)>ASU1_(b))?a:b;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 AMinD1(AD1 a,AD1 b){return a>ASL1_(b));} + A_STATIC AU1 AShrSU1(AU1 a,AU1 b){return AU1_(ASU1_(a)>>ASU1_(b));} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ASinD1(AD1 a){return __builtin_sin(a);} + A_STATIC AF1 ASinF1(AF1 a){return __builtin_sinf(a);} + #else + A_STATIC AD1 ASinD1(AD1 a){return sin(a);} + A_STATIC AF1 ASinF1(AF1 a){return sinf(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ASqrtD1(AD1 a){return __builtin_sqrt(a);} + A_STATIC AF1 ASqrtF1(AF1 a){return __builtin_sqrtf(a);} + #else + A_STATIC AD1 ASqrtD1(AD1 a){return sqrt(a);} + A_STATIC AF1 ASqrtF1(AF1 a){return sqrtf(a);} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS - DEPENDENT +//============================================================================================================================== + A_STATIC AD1 AClampD1(AD1 x,AD1 n,AD1 m){return AMaxD1(n,AMinD1(x,m));} + A_STATIC AF1 AClampF1(AF1 x,AF1 n,AF1 m){return AMaxF1(n,AMinF1(x,m));} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 AFractD1(AD1 a){return a-AFloorD1(a);} + A_STATIC AF1 AFractF1(AF1 a){return a-AFloorF1(a);} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 APowD1(AD1 a,AD1 b){return AExp2D1(b*ALog2D1(a));} + A_STATIC AF1 APowF1(AF1 a,AF1 b){return AExp2F1(b*ALog2F1(a));} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ARsqD1(AD1 a){return ARcpD1(ASqrtD1(a));} + A_STATIC AF1 ARsqF1(AF1 a){return ARcpF1(ASqrtF1(a));} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ASatD1(AD1 a){return AMinD1(1.0,AMaxD1(0.0,a));} + A_STATIC AF1 ASatF1(AF1 a){return AMinF1(1.0f,AMaxF1(0.0f,a));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR OPS +//------------------------------------------------------------------------------------------------------------------------------ +// These are added as needed for production or prototyping, so not necessarily a complete set. +// They follow a convention of taking in a destination and also returning the destination value to increase utility. +//============================================================================================================================== + A_STATIC retAD2 opAAbsD2(outAD2 d,inAD2 a){d[0]=AAbsD1(a[0]);d[1]=AAbsD1(a[1]);return d;} + A_STATIC retAD3 opAAbsD3(outAD3 d,inAD3 a){d[0]=AAbsD1(a[0]);d[1]=AAbsD1(a[1]);d[2]=AAbsD1(a[2]);return d;} + A_STATIC retAD4 opAAbsD4(outAD4 d,inAD4 a){d[0]=AAbsD1(a[0]);d[1]=AAbsD1(a[1]);d[2]=AAbsD1(a[2]);d[3]=AAbsD1(a[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAAbsF2(outAF2 d,inAF2 a){d[0]=AAbsF1(a[0]);d[1]=AAbsF1(a[1]);return d;} + A_STATIC retAF3 opAAbsF3(outAF3 d,inAF3 a){d[0]=AAbsF1(a[0]);d[1]=AAbsF1(a[1]);d[2]=AAbsF1(a[2]);return d;} + A_STATIC retAF4 opAAbsF4(outAF4 d,inAF4 a){d[0]=AAbsF1(a[0]);d[1]=AAbsF1(a[1]);d[2]=AAbsF1(a[2]);d[3]=AAbsF1(a[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAAddD2(outAD2 d,inAD2 a,inAD2 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];return d;} + A_STATIC retAD3 opAAddD3(outAD3 d,inAD3 a,inAD3 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];return d;} + A_STATIC retAD4 opAAddD4(outAD4 d,inAD4 a,inAD4 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];d[3]=a[3]+b[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAAddF2(outAF2 d,inAF2 a,inAF2 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];return d;} + A_STATIC retAF3 opAAddF3(outAF3 d,inAF3 a,inAF3 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];return d;} + A_STATIC retAF4 opAAddF4(outAF4 d,inAF4 a,inAF4 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];d[3]=a[3]+b[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opAAddOneD2(outAD2 d,inAD2 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;return d;} + A_STATIC retAD3 opAAddOneD3(outAD3 d,inAD3 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;return d;} + A_STATIC retAD4 opAAddOneD4(outAD4 d,inAD4 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;d[3]=a[3]+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAAddOneF2(outAF2 d,inAF2 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;return d;} + A_STATIC retAF3 opAAddOneF3(outAF3 d,inAF3 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;return d;} + A_STATIC retAF4 opAAddOneF4(outAF4 d,inAF4 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;d[3]=a[3]+b;return d;} +//============================================================================================================================== + A_STATIC retAD2 opACpyD2(outAD2 d,inAD2 a){d[0]=a[0];d[1]=a[1];return d;} + A_STATIC retAD3 opACpyD3(outAD3 d,inAD3 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];return d;} + A_STATIC retAD4 opACpyD4(outAD4 d,inAD4 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];d[3]=a[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opACpyF2(outAF2 d,inAF2 a){d[0]=a[0];d[1]=a[1];return d;} + A_STATIC retAF3 opACpyF3(outAF3 d,inAF3 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];return d;} + A_STATIC retAF4 opACpyF4(outAF4 d,inAF4 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];d[3]=a[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opALerpD2(outAD2 d,inAD2 a,inAD2 b,inAD2 c){d[0]=ALerpD1(a[0],b[0],c[0]);d[1]=ALerpD1(a[1],b[1],c[1]);return d;} + A_STATIC retAD3 opALerpD3(outAD3 d,inAD3 a,inAD3 b,inAD3 c){d[0]=ALerpD1(a[0],b[0],c[0]);d[1]=ALerpD1(a[1],b[1],c[1]);d[2]=ALerpD1(a[2],b[2],c[2]);return d;} + A_STATIC retAD4 opALerpD4(outAD4 d,inAD4 a,inAD4 b,inAD4 c){d[0]=ALerpD1(a[0],b[0],c[0]);d[1]=ALerpD1(a[1],b[1],c[1]);d[2]=ALerpD1(a[2],b[2],c[2]);d[3]=ALerpD1(a[3],b[3],c[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opALerpF2(outAF2 d,inAF2 a,inAF2 b,inAF2 c){d[0]=ALerpF1(a[0],b[0],c[0]);d[1]=ALerpF1(a[1],b[1],c[1]);return d;} + A_STATIC retAF3 opALerpF3(outAF3 d,inAF3 a,inAF3 b,inAF3 c){d[0]=ALerpF1(a[0],b[0],c[0]);d[1]=ALerpF1(a[1],b[1],c[1]);d[2]=ALerpF1(a[2],b[2],c[2]);return d;} + A_STATIC retAF4 opALerpF4(outAF4 d,inAF4 a,inAF4 b,inAF4 c){d[0]=ALerpF1(a[0],b[0],c[0]);d[1]=ALerpF1(a[1],b[1],c[1]);d[2]=ALerpF1(a[2],b[2],c[2]);d[3]=ALerpF1(a[3],b[3],c[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opALerpOneD2(outAD2 d,inAD2 a,inAD2 b,AD1 c){d[0]=ALerpD1(a[0],b[0],c);d[1]=ALerpD1(a[1],b[1],c);return d;} + A_STATIC retAD3 opALerpOneD3(outAD3 d,inAD3 a,inAD3 b,AD1 c){d[0]=ALerpD1(a[0],b[0],c);d[1]=ALerpD1(a[1],b[1],c);d[2]=ALerpD1(a[2],b[2],c);return d;} + A_STATIC retAD4 opALerpOneD4(outAD4 d,inAD4 a,inAD4 b,AD1 c){d[0]=ALerpD1(a[0],b[0],c);d[1]=ALerpD1(a[1],b[1],c);d[2]=ALerpD1(a[2],b[2],c);d[3]=ALerpD1(a[3],b[3],c);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opALerpOneF2(outAF2 d,inAF2 a,inAF2 b,AF1 c){d[0]=ALerpF1(a[0],b[0],c);d[1]=ALerpF1(a[1],b[1],c);return d;} + A_STATIC retAF3 opALerpOneF3(outAF3 d,inAF3 a,inAF3 b,AF1 c){d[0]=ALerpF1(a[0],b[0],c);d[1]=ALerpF1(a[1],b[1],c);d[2]=ALerpF1(a[2],b[2],c);return d;} + A_STATIC retAF4 opALerpOneF4(outAF4 d,inAF4 a,inAF4 b,AF1 c){d[0]=ALerpF1(a[0],b[0],c);d[1]=ALerpF1(a[1],b[1],c);d[2]=ALerpF1(a[2],b[2],c);d[3]=ALerpF1(a[3],b[3],c);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMaxD2(outAD2 d,inAD2 a,inAD2 b){d[0]=AMaxD1(a[0],b[0]);d[1]=AMaxD1(a[1],b[1]);return d;} + A_STATIC retAD3 opAMaxD3(outAD3 d,inAD3 a,inAD3 b){d[0]=AMaxD1(a[0],b[0]);d[1]=AMaxD1(a[1],b[1]);d[2]=AMaxD1(a[2],b[2]);return d;} + A_STATIC retAD4 opAMaxD4(outAD4 d,inAD4 a,inAD4 b){d[0]=AMaxD1(a[0],b[0]);d[1]=AMaxD1(a[1],b[1]);d[2]=AMaxD1(a[2],b[2]);d[3]=AMaxD1(a[3],b[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMaxF2(outAF2 d,inAF2 a,inAF2 b){d[0]=AMaxF1(a[0],b[0]);d[1]=AMaxF1(a[1],b[1]);return d;} + A_STATIC retAF3 opAMaxF3(outAF3 d,inAF3 a,inAF3 b){d[0]=AMaxF1(a[0],b[0]);d[1]=AMaxF1(a[1],b[1]);d[2]=AMaxF1(a[2],b[2]);return d;} + A_STATIC retAF4 opAMaxF4(outAF4 d,inAF4 a,inAF4 b){d[0]=AMaxF1(a[0],b[0]);d[1]=AMaxF1(a[1],b[1]);d[2]=AMaxF1(a[2],b[2]);d[3]=AMaxF1(a[3],b[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMinD2(outAD2 d,inAD2 a,inAD2 b){d[0]=AMinD1(a[0],b[0]);d[1]=AMinD1(a[1],b[1]);return d;} + A_STATIC retAD3 opAMinD3(outAD3 d,inAD3 a,inAD3 b){d[0]=AMinD1(a[0],b[0]);d[1]=AMinD1(a[1],b[1]);d[2]=AMinD1(a[2],b[2]);return d;} + A_STATIC retAD4 opAMinD4(outAD4 d,inAD4 a,inAD4 b){d[0]=AMinD1(a[0],b[0]);d[1]=AMinD1(a[1],b[1]);d[2]=AMinD1(a[2],b[2]);d[3]=AMinD1(a[3],b[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMinF2(outAF2 d,inAF2 a,inAF2 b){d[0]=AMinF1(a[0],b[0]);d[1]=AMinF1(a[1],b[1]);return d;} + A_STATIC retAF3 opAMinF3(outAF3 d,inAF3 a,inAF3 b){d[0]=AMinF1(a[0],b[0]);d[1]=AMinF1(a[1],b[1]);d[2]=AMinF1(a[2],b[2]);return d;} + A_STATIC retAF4 opAMinF4(outAF4 d,inAF4 a,inAF4 b){d[0]=AMinF1(a[0],b[0]);d[1]=AMinF1(a[1],b[1]);d[2]=AMinF1(a[2],b[2]);d[3]=AMinF1(a[3],b[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMulD2(outAD2 d,inAD2 a,inAD2 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];return d;} + A_STATIC retAD3 opAMulD3(outAD3 d,inAD3 a,inAD3 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];return d;} + A_STATIC retAD4 opAMulD4(outAD4 d,inAD4 a,inAD4 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];d[3]=a[3]*b[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMulF2(outAF2 d,inAF2 a,inAF2 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];return d;} + A_STATIC retAF3 opAMulF3(outAF3 d,inAF3 a,inAF3 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];return d;} + A_STATIC retAF4 opAMulF4(outAF4 d,inAF4 a,inAF4 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];d[3]=a[3]*b[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMulOneD2(outAD2 d,inAD2 a,AD1 b){d[0]=a[0]*b;d[1]=a[1]*b;return d;} + A_STATIC retAD3 opAMulOneD3(outAD3 d,inAD3 a,AD1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;return d;} + A_STATIC retAD4 opAMulOneD4(outAD4 d,inAD4 a,AD1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;d[3]=a[3]*b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMulOneF2(outAF2 d,inAF2 a,AF1 b){d[0]=a[0]*b;d[1]=a[1]*b;return d;} + A_STATIC retAF3 opAMulOneF3(outAF3 d,inAF3 a,AF1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;return d;} + A_STATIC retAF4 opAMulOneF4(outAF4 d,inAF4 a,AF1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;d[3]=a[3]*b;return d;} +//============================================================================================================================== + A_STATIC retAD2 opANegD2(outAD2 d,inAD2 a){d[0]=-a[0];d[1]=-a[1];return d;} + A_STATIC retAD3 opANegD3(outAD3 d,inAD3 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];return d;} + A_STATIC retAD4 opANegD4(outAD4 d,inAD4 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];d[3]=-a[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opANegF2(outAF2 d,inAF2 a){d[0]=-a[0];d[1]=-a[1];return d;} + A_STATIC retAF3 opANegF3(outAF3 d,inAF3 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];return d;} + A_STATIC retAF4 opANegF4(outAF4 d,inAF4 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];d[3]=-a[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opARcpD2(outAD2 d,inAD2 a){d[0]=ARcpD1(a[0]);d[1]=ARcpD1(a[1]);return d;} + A_STATIC retAD3 opARcpD3(outAD3 d,inAD3 a){d[0]=ARcpD1(a[0]);d[1]=ARcpD1(a[1]);d[2]=ARcpD1(a[2]);return d;} + A_STATIC retAD4 opARcpD4(outAD4 d,inAD4 a){d[0]=ARcpD1(a[0]);d[1]=ARcpD1(a[1]);d[2]=ARcpD1(a[2]);d[3]=ARcpD1(a[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opARcpF2(outAF2 d,inAF2 a){d[0]=ARcpF1(a[0]);d[1]=ARcpF1(a[1]);return d;} + A_STATIC retAF3 opARcpF3(outAF3 d,inAF3 a){d[0]=ARcpF1(a[0]);d[1]=ARcpF1(a[1]);d[2]=ARcpF1(a[2]);return d;} + A_STATIC retAF4 opARcpF4(outAF4 d,inAF4 a){d[0]=ARcpF1(a[0]);d[1]=ARcpF1(a[1]);d[2]=ARcpF1(a[2]);d[3]=ARcpF1(a[3]);return d;} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HALF FLOAT PACKING +//============================================================================================================================== + // Convert float to half (in lower 16-bits of output). + // Same fast technique as documented here: ftp://ftp.fox-toolkit.org/pub/fasthalffloatconversion.pdf + // Supports denormals. + // Conversion rules are to make computations possibly "safer" on the GPU, + // -INF & -NaN -> -65504 + // +INF & +NaN -> +65504 + A_STATIC AU1 AU1_AH1_AF1(AF1 f){ + static AW1 base[512]={ + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0001,0x0002,0x0004,0x0008,0x0010,0x0020,0x0040,0x0080,0x0100, + 0x0200,0x0400,0x0800,0x0c00,0x1000,0x1400,0x1800,0x1c00,0x2000,0x2400,0x2800,0x2c00,0x3000,0x3400,0x3800,0x3c00, + 0x4000,0x4400,0x4800,0x4c00,0x5000,0x5400,0x5800,0x5c00,0x6000,0x6400,0x6800,0x6c00,0x7000,0x7400,0x7800,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8001,0x8002,0x8004,0x8008,0x8010,0x8020,0x8040,0x8080,0x8100, + 0x8200,0x8400,0x8800,0x8c00,0x9000,0x9400,0x9800,0x9c00,0xa000,0xa400,0xa800,0xac00,0xb000,0xb400,0xb800,0xbc00, + 0xc000,0xc400,0xc800,0xcc00,0xd000,0xd400,0xd800,0xdc00,0xe000,0xe400,0xe800,0xec00,0xf000,0xf400,0xf800,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff}; + static AB1 shift[512]={ + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x17,0x16,0x15,0x14,0x13,0x12,0x11,0x10,0x0f, + 0x0e,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d, + 0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x17,0x16,0x15,0x14,0x13,0x12,0x11,0x10,0x0f, + 0x0e,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d, + 0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18}; + union{AF1 f;AU1 u;}bits;bits.f=f;AU1 u=bits.u;AU1 i=u>>23;return (AU1)(base[i])+((u&0x7fffff)>>shift[i]);} +//------------------------------------------------------------------------------------------------------------------------------ + // Used to output packed constant. + A_STATIC AU1 AU1_AH2_AF2(inAF2 a){return AU1_AH1_AF1(a[0])+(AU1_AH1_AF1(a[1])<<16);} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// GLSL +// +// +//============================================================================================================================== +#if defined(A_GLSL) && defined(A_GPU) + #ifndef A_SKIP_EXT + #ifdef A_HALF + #extension GL_EXT_shader_16bit_storage:require + #extension GL_EXT_shader_explicit_arithmetic_types:require + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_LONG + #extension GL_ARB_gpu_shader_int64:require + #extension GL_NV_shader_atomic_int64:require + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_WAVE + #extension GL_KHR_shader_subgroup_arithmetic:require + #extension GL_KHR_shader_subgroup_ballot:require + #extension GL_KHR_shader_subgroup_quad:require + #extension GL_KHR_shader_subgroup_shuffle:require + #endif + #endif +//============================================================================================================================== + #define AP1 bool + #define AP2 bvec2 + #define AP3 bvec3 + #define AP4 bvec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF1 float + #define AF2 vec2 + #define AF3 vec3 + #define AF4 vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1 uint + #define AU2 uvec2 + #define AU3 uvec3 + #define AU4 uvec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASU1 int + #define ASU2 ivec2 + #define ASU3 ivec3 + #define ASU4 ivec4 +//============================================================================================================================== + #define AF1_AU1(x) uintBitsToFloat(AU1(x)) + #define AF2_AU2(x) uintBitsToFloat(AU2(x)) + #define AF3_AU3(x) uintBitsToFloat(AU3(x)) + #define AF4_AU4(x) uintBitsToFloat(AU4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AF1(x) floatBitsToUint(AF1(x)) + #define AU2_AF2(x) floatBitsToUint(AF2(x)) + #define AU3_AF3(x) floatBitsToUint(AF3(x)) + #define AU4_AF4(x) floatBitsToUint(AF4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH1_AF1_x(AF1 a){return packHalf2x16(AF2(a,0.0));} + #define AU1_AH1_AF1(a) AU1_AH1_AF1_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AH2_AF2 packHalf2x16 + #define AU1_AW2Unorm_AF2 packUnorm2x16 + #define AU1_AB4Unorm_AF4 packUnorm4x8 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF2_AH2_AU1 unpackHalf2x16 + #define AF2_AW2Unorm_AU1 unpackUnorm2x16 + #define AF4_AB4Unorm_AU1 unpackUnorm4x8 +//============================================================================================================================== + AF1 AF1_x(AF1 a){return AF1(a);} + AF2 AF2_x(AF1 a){return AF2(a,a);} + AF3 AF3_x(AF1 a){return AF3(a,a,a);} + AF4 AF4_x(AF1 a){return AF4(a,a,a,a);} + #define AF1_(a) AF1_x(AF1(a)) + #define AF2_(a) AF2_x(AF1(a)) + #define AF3_(a) AF3_x(AF1(a)) + #define AF4_(a) AF4_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_x(AU1 a){return AU1(a);} + AU2 AU2_x(AU1 a){return AU2(a,a);} + AU3 AU3_x(AU1 a){return AU3(a,a,a);} + AU4 AU4_x(AU1 a){return AU4(a,a,a,a);} + #define AU1_(a) AU1_x(AU1(a)) + #define AU2_(a) AU2_x(AU1(a)) + #define AU3_(a) AU3_x(AU1(a)) + #define AU4_(a) AU4_x(AU1(a)) +//============================================================================================================================== + AU1 AAbsSU1(AU1 a){return AU1(abs(ASU1(a)));} + AU2 AAbsSU2(AU2 a){return AU2(abs(ASU2(a)));} + AU3 AAbsSU3(AU3 a){return AU3(abs(ASU3(a)));} + AU4 AAbsSU4(AU4 a){return AU4(abs(ASU4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 ABfe(AU1 src,AU1 off,AU1 bits){return bitfieldExtract(src,ASU1(off),ASU1(bits));} + AU1 ABfi(AU1 src,AU1 ins,AU1 mask){return (ins&mask)|(src&(~mask));} + // Proxy for V_BFI_B32 where the 'mask' is set as 'bits', 'mask=(1<>ASU1(b));} + AU2 AShrSU2(AU2 a,AU2 b){return AU2(ASU2(a)>>ASU2(b));} + AU3 AShrSU3(AU3 a,AU3 b){return AU3(ASU3(a)>>ASU3(b));} + AU4 AShrSU4(AU4 a,AU4 b){return AU4(ASU4(a)>>ASU4(b));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL BYTE +//============================================================================================================================== + #ifdef A_BYTE + #define AB1 uint8_t + #define AB2 u8vec2 + #define AB3 u8vec3 + #define AB4 u8vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASB1 int8_t + #define ASB2 i8vec2 + #define ASB3 i8vec3 + #define ASB4 i8vec4 +//------------------------------------------------------------------------------------------------------------------------------ + AB1 AB1_x(AB1 a){return AB1(a);} + AB2 AB2_x(AB1 a){return AB2(a,a);} + AB3 AB3_x(AB1 a){return AB3(a,a,a);} + AB4 AB4_x(AB1 a){return AB4(a,a,a,a);} + #define AB1_(a) AB1_x(AB1(a)) + #define AB2_(a) AB2_x(AB1(a)) + #define AB3_(a) AB3_x(AB1(a)) + #define AB4_(a) AB4_x(AB1(a)) + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL HALF +//============================================================================================================================== + #ifdef A_HALF + #define AH1 float16_t + #define AH2 f16vec2 + #define AH3 f16vec3 + #define AH4 f16vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AW1 uint16_t + #define AW2 u16vec2 + #define AW3 u16vec3 + #define AW4 u16vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASW1 int16_t + #define ASW2 i16vec2 + #define ASW3 i16vec3 + #define ASW4 i16vec4 +//============================================================================================================================== + #define AH2_AU1(x) unpackFloat2x16(AU1(x)) + AH4 AH4_AU2_x(AU2 x){return AH4(unpackFloat2x16(x.x),unpackFloat2x16(x.y));} + #define AH4_AU2(x) AH4_AU2_x(AU2(x)) + #define AW2_AU1(x) unpackUint2x16(AU1(x)) + #define AW4_AU2(x) unpackUint4x16(pack64(AU2(x))) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AH2(x) packFloat2x16(AH2(x)) + AU2 AU2_AH4_x(AH4 x){return AU2(packFloat2x16(x.xy),packFloat2x16(x.zw));} + #define AU2_AH4(x) AU2_AH4_x(AH4(x)) + #define AU1_AW2(x) packUint2x16(AW2(x)) + #define AU2_AW4(x) unpack32(packUint4x16(AW4(x))) +//============================================================================================================================== + #define AW1_AH1(x) halfBitsToUint16(AH1(x)) + #define AW2_AH2(x) halfBitsToUint16(AH2(x)) + #define AW3_AH3(x) halfBitsToUint16(AH3(x)) + #define AW4_AH4(x) halfBitsToUint16(AH4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AH1_AW1(x) uint16BitsToHalf(AW1(x)) + #define AH2_AW2(x) uint16BitsToHalf(AW2(x)) + #define AH3_AW3(x) uint16BitsToHalf(AW3(x)) + #define AH4_AW4(x) uint16BitsToHalf(AW4(x)) +//============================================================================================================================== + AH1 AH1_x(AH1 a){return AH1(a);} + AH2 AH2_x(AH1 a){return AH2(a,a);} + AH3 AH3_x(AH1 a){return AH3(a,a,a);} + AH4 AH4_x(AH1 a){return AH4(a,a,a,a);} + #define AH1_(a) AH1_x(AH1(a)) + #define AH2_(a) AH2_x(AH1(a)) + #define AH3_(a) AH3_x(AH1(a)) + #define AH4_(a) AH4_x(AH1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AW1_x(AW1 a){return AW1(a);} + AW2 AW2_x(AW1 a){return AW2(a,a);} + AW3 AW3_x(AW1 a){return AW3(a,a,a);} + AW4 AW4_x(AW1 a){return AW4(a,a,a,a);} + #define AW1_(a) AW1_x(AW1(a)) + #define AW2_(a) AW2_x(AW1(a)) + #define AW3_(a) AW3_x(AW1(a)) + #define AW4_(a) AW4_x(AW1(a)) +//============================================================================================================================== + AW1 AAbsSW1(AW1 a){return AW1(abs(ASW1(a)));} + AW2 AAbsSW2(AW2 a){return AW2(abs(ASW2(a)));} + AW3 AAbsSW3(AW3 a){return AW3(abs(ASW3(a)));} + AW4 AAbsSW4(AW4 a){return AW4(abs(ASW4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AClampH1(AH1 x,AH1 n,AH1 m){return clamp(x,n,m);} + AH2 AClampH2(AH2 x,AH2 n,AH2 m){return clamp(x,n,m);} + AH3 AClampH3(AH3 x,AH3 n,AH3 m){return clamp(x,n,m);} + AH4 AClampH4(AH4 x,AH4 n,AH4 m){return clamp(x,n,m);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFractH1(AH1 x){return fract(x);} + AH2 AFractH2(AH2 x){return fract(x);} + AH3 AFractH3(AH3 x){return fract(x);} + AH4 AFractH4(AH4 x){return fract(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ALerpH1(AH1 x,AH1 y,AH1 a){return mix(x,y,a);} + AH2 ALerpH2(AH2 x,AH2 y,AH2 a){return mix(x,y,a);} + AH3 ALerpH3(AH3 x,AH3 y,AH3 a){return mix(x,y,a);} + AH4 ALerpH4(AH4 x,AH4 y,AH4 a){return mix(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + // No packed version of max3. + AH1 AMax3H1(AH1 x,AH1 y,AH1 z){return max(x,max(y,z));} + AH2 AMax3H2(AH2 x,AH2 y,AH2 z){return max(x,max(y,z));} + AH3 AMax3H3(AH3 x,AH3 y,AH3 z){return max(x,max(y,z));} + AH4 AMax3H4(AH4 x,AH4 y,AH4 z){return max(x,max(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMaxSW1(AW1 a,AW1 b){return AW1(max(ASU1(a),ASU1(b)));} + AW2 AMaxSW2(AW2 a,AW2 b){return AW2(max(ASU2(a),ASU2(b)));} + AW3 AMaxSW3(AW3 a,AW3 b){return AW3(max(ASU3(a),ASU3(b)));} + AW4 AMaxSW4(AW4 a,AW4 b){return AW4(max(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + // No packed version of min3. + AH1 AMin3H1(AH1 x,AH1 y,AH1 z){return min(x,min(y,z));} + AH2 AMin3H2(AH2 x,AH2 y,AH2 z){return min(x,min(y,z));} + AH3 AMin3H3(AH3 x,AH3 y,AH3 z){return min(x,min(y,z));} + AH4 AMin3H4(AH4 x,AH4 y,AH4 z){return min(x,min(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMinSW1(AW1 a,AW1 b){return AW1(min(ASU1(a),ASU1(b)));} + AW2 AMinSW2(AW2 a,AW2 b){return AW2(min(ASU2(a),ASU2(b)));} + AW3 AMinSW3(AW3 a,AW3 b){return AW3(min(ASU3(a),ASU3(b)));} + AW4 AMinSW4(AW4 a,AW4 b){return AW4(min(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARcpH1(AH1 x){return AH1_(1.0)/x;} + AH2 ARcpH2(AH2 x){return AH2_(1.0)/x;} + AH3 ARcpH3(AH3 x){return AH3_(1.0)/x;} + AH4 ARcpH4(AH4 x){return AH4_(1.0)/x;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARsqH1(AH1 x){return AH1_(1.0)/sqrt(x);} + AH2 ARsqH2(AH2 x){return AH2_(1.0)/sqrt(x);} + AH3 ARsqH3(AH3 x){return AH3_(1.0)/sqrt(x);} + AH4 ARsqH4(AH4 x){return AH4_(1.0)/sqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASatH1(AH1 x){return clamp(x,AH1_(0.0),AH1_(1.0));} + AH2 ASatH2(AH2 x){return clamp(x,AH2_(0.0),AH2_(1.0));} + AH3 ASatH3(AH3 x){return clamp(x,AH3_(0.0),AH3_(1.0));} + AH4 ASatH4(AH4 x){return clamp(x,AH4_(0.0),AH4_(1.0));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AShrSW1(AW1 a,AW1 b){return AW1(ASW1(a)>>ASW1(b));} + AW2 AShrSW2(AW2 a,AW2 b){return AW2(ASW2(a)>>ASW2(b));} + AW3 AShrSW3(AW3 a,AW3 b){return AW3(ASW3(a)>>ASW3(b));} + AW4 AShrSW4(AW4 a,AW4 b){return AW4(ASW4(a)>>ASW4(b));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL DOUBLE +//============================================================================================================================== + #ifdef A_DUBL + #define AD1 double + #define AD2 dvec2 + #define AD3 dvec3 + #define AD4 dvec4 +//------------------------------------------------------------------------------------------------------------------------------ + AD1 AD1_x(AD1 a){return AD1(a);} + AD2 AD2_x(AD1 a){return AD2(a,a);} + AD3 AD3_x(AD1 a){return AD3(a,a,a);} + AD4 AD4_x(AD1 a){return AD4(a,a,a,a);} + #define AD1_(a) AD1_x(AD1(a)) + #define AD2_(a) AD2_x(AD1(a)) + #define AD3_(a) AD3_x(AD1(a)) + #define AD4_(a) AD4_x(AD1(a)) +//============================================================================================================================== + AD1 AFractD1(AD1 x){return fract(x);} + AD2 AFractD2(AD2 x){return fract(x);} + AD3 AFractD3(AD3 x){return fract(x);} + AD4 AFractD4(AD4 x){return fract(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ALerpD1(AD1 x,AD1 y,AD1 a){return mix(x,y,a);} + AD2 ALerpD2(AD2 x,AD2 y,AD2 a){return mix(x,y,a);} + AD3 ALerpD3(AD3 x,AD3 y,AD3 a){return mix(x,y,a);} + AD4 ALerpD4(AD4 x,AD4 y,AD4 a){return mix(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARcpD1(AD1 x){return AD1_(1.0)/x;} + AD2 ARcpD2(AD2 x){return AD2_(1.0)/x;} + AD3 ARcpD3(AD3 x){return AD3_(1.0)/x;} + AD4 ARcpD4(AD4 x){return AD4_(1.0)/x;} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARsqD1(AD1 x){return AD1_(1.0)/sqrt(x);} + AD2 ARsqD2(AD2 x){return AD2_(1.0)/sqrt(x);} + AD3 ARsqD3(AD3 x){return AD3_(1.0)/sqrt(x);} + AD4 ARsqD4(AD4 x){return AD4_(1.0)/sqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ASatD1(AD1 x){return clamp(x,AD1_(0.0),AD1_(1.0));} + AD2 ASatD2(AD2 x){return clamp(x,AD2_(0.0),AD2_(1.0));} + AD3 ASatD3(AD3 x){return clamp(x,AD3_(0.0),AD3_(1.0));} + AD4 ASatD4(AD4 x){return clamp(x,AD4_(0.0),AD4_(1.0));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL LONG +//============================================================================================================================== + #ifdef A_LONG + #define AL1 uint64_t + #define AL2 u64vec2 + #define AL3 u64vec3 + #define AL4 u64vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASL1 int64_t + #define ASL2 i64vec2 + #define ASL3 i64vec3 + #define ASL4 i64vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AL1_AU2(x) packUint2x32(AU2(x)) + #define AU2_AL1(x) unpackUint2x32(AL1(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AL1 AL1_x(AL1 a){return AL1(a);} + AL2 AL2_x(AL1 a){return AL2(a,a);} + AL3 AL3_x(AL1 a){return AL3(a,a,a);} + AL4 AL4_x(AL1 a){return AL4(a,a,a,a);} + #define AL1_(a) AL1_x(AL1(a)) + #define AL2_(a) AL2_x(AL1(a)) + #define AL3_(a) AL3_x(AL1(a)) + #define AL4_(a) AL4_x(AL1(a)) +//============================================================================================================================== + AL1 AAbsSL1(AL1 a){return AL1(abs(ASL1(a)));} + AL2 AAbsSL2(AL2 a){return AL2(abs(ASL2(a)));} + AL3 AAbsSL3(AL3 a){return AL3(abs(ASL3(a)));} + AL4 AAbsSL4(AL4 a){return AL4(abs(ASL4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AL1 AMaxSL1(AL1 a,AL1 b){return AL1(max(ASU1(a),ASU1(b)));} + AL2 AMaxSL2(AL2 a,AL2 b){return AL2(max(ASU2(a),ASU2(b)));} + AL3 AMaxSL3(AL3 a,AL3 b){return AL3(max(ASU3(a),ASU3(b)));} + AL4 AMaxSL4(AL4 a,AL4 b){return AL4(max(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AL1 AMinSL1(AL1 a,AL1 b){return AL1(min(ASU1(a),ASU1(b)));} + AL2 AMinSL2(AL2 a,AL2 b){return AL2(min(ASU2(a),ASU2(b)));} + AL3 AMinSL3(AL3 a,AL3 b){return AL3(min(ASU3(a),ASU3(b)));} + AL4 AMinSL4(AL4 a,AL4 b){return AL4(min(ASU4(a),ASU4(b)));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// WAVE OPERATIONS +//============================================================================================================================== + #ifdef A_WAVE + // Where 'x' must be a compile time literal. + AF1 AWaveXorF1(AF1 v,AU1 x){return subgroupShuffleXor(v,x);} + AF2 AWaveXorF2(AF2 v,AU1 x){return subgroupShuffleXor(v,x);} + AF3 AWaveXorF3(AF3 v,AU1 x){return subgroupShuffleXor(v,x);} + AF4 AWaveXorF4(AF4 v,AU1 x){return subgroupShuffleXor(v,x);} + AU1 AWaveXorU1(AU1 v,AU1 x){return subgroupShuffleXor(v,x);} + AU2 AWaveXorU2(AU2 v,AU1 x){return subgroupShuffleXor(v,x);} + AU3 AWaveXorU3(AU3 v,AU1 x){return subgroupShuffleXor(v,x);} + AU4 AWaveXorU4(AU4 v,AU1 x){return subgroupShuffleXor(v,x);} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + AH2 AWaveXorH2(AH2 v,AU1 x){return AH2_AU1(subgroupShuffleXor(AU1_AH2(v),x));} + AH4 AWaveXorH4(AH4 v,AU1 x){return AH4_AU2(subgroupShuffleXor(AU2_AH4(v),x));} + AW2 AWaveXorW2(AW2 v,AU1 x){return AW2_AU1(subgroupShuffleXor(AU1_AW2(v),x));} + AW4 AWaveXorW4(AW4 v,AU1 x){return AW4_AU2(subgroupShuffleXor(AU2_AW4(v),x));} + #endif + #endif +//============================================================================================================================== +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// HLSL +// +// +//============================================================================================================================== +#if defined(A_HLSL) && defined(A_GPU) + #ifdef A_HLSL_6_2 + #define AP1 bool + #define AP2 bool2 + #define AP3 bool3 + #define AP4 bool4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF1 float32_t + #define AF2 float32_t2 + #define AF3 float32_t3 + #define AF4 float32_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1 uint32_t + #define AU2 uint32_t2 + #define AU3 uint32_t3 + #define AU4 uint32_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASU1 int32_t + #define ASU2 int32_t2 + #define ASU3 int32_t3 + #define ASU4 int32_t4 + #else + #define AP1 bool + #define AP2 bool2 + #define AP3 bool3 + #define AP4 bool4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF1 float + #define AF2 float2 + #define AF3 float3 + #define AF4 float4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1 uint + #define AU2 uint2 + #define AU3 uint3 + #define AU4 uint4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASU1 int + #define ASU2 int2 + #define ASU3 int3 + #define ASU4 int4 + #endif +//============================================================================================================================== + #define AF1_AU1(x) asfloat(AU1(x)) + #define AF2_AU2(x) asfloat(AU2(x)) + #define AF3_AU3(x) asfloat(AU3(x)) + #define AF4_AU4(x) asfloat(AU4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AF1(x) asuint(AF1(x)) + #define AU2_AF2(x) asuint(AF2(x)) + #define AU3_AF3(x) asuint(AF3(x)) + #define AU4_AF4(x) asuint(AF4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH1_AF1_x(AF1 a){return f32tof16(a);} + #define AU1_AH1_AF1(a) AU1_AH1_AF1_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH2_AF2_x(AF2 a){return f32tof16(a.x)|(f32tof16(a.y)<<16);} + #define AU1_AH2_AF2(a) AU1_AH2_AF2_x(AF2(a)) + #define AU1_AB4Unorm_AF4(x) D3DCOLORtoUBYTE4(AF4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AF2 AF2_AH2_AU1_x(AU1 x){return AF2(f16tof32(x&0xFFFF),f16tof32(x>>16));} + #define AF2_AH2_AU1(x) AF2_AH2_AU1_x(AU1(x)) +//============================================================================================================================== + AF1 AF1_x(AF1 a){return AF1(a);} + AF2 AF2_x(AF1 a){return AF2(a,a);} + AF3 AF3_x(AF1 a){return AF3(a,a,a);} + AF4 AF4_x(AF1 a){return AF4(a,a,a,a);} + #define AF1_(a) AF1_x(AF1(a)) + #define AF2_(a) AF2_x(AF1(a)) + #define AF3_(a) AF3_x(AF1(a)) + #define AF4_(a) AF4_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_x(AU1 a){return AU1(a);} + AU2 AU2_x(AU1 a){return AU2(a,a);} + AU3 AU3_x(AU1 a){return AU3(a,a,a);} + AU4 AU4_x(AU1 a){return AU4(a,a,a,a);} + #define AU1_(a) AU1_x(AU1(a)) + #define AU2_(a) AU2_x(AU1(a)) + #define AU3_(a) AU3_x(AU1(a)) + #define AU4_(a) AU4_x(AU1(a)) +//============================================================================================================================== + AU1 AAbsSU1(AU1 a){return AU1(abs(ASU1(a)));} + AU2 AAbsSU2(AU2 a){return AU2(abs(ASU2(a)));} + AU3 AAbsSU3(AU3 a){return AU3(abs(ASU3(a)));} + AU4 AAbsSU4(AU4 a){return AU4(abs(ASU4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 ABfe(AU1 src,AU1 off,AU1 bits){AU1 mask=(1u<>off)&mask;} + AU1 ABfi(AU1 src,AU1 ins,AU1 mask){return (ins&mask)|(src&(~mask));} + AU1 ABfiM(AU1 src,AU1 ins,AU1 bits){AU1 mask=(1u<>ASU1(b));} + AU2 AShrSU2(AU2 a,AU2 b){return AU2(ASU2(a)>>ASU2(b));} + AU3 AShrSU3(AU3 a,AU3 b){return AU3(ASU3(a)>>ASU3(b));} + AU4 AShrSU4(AU4 a,AU4 b){return AU4(ASU4(a)>>ASU4(b));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HLSL BYTE +//============================================================================================================================== + #ifdef A_BYTE + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HLSL HALF +//============================================================================================================================== + #ifdef A_HALF + #ifdef A_HLSL_6_2 + #define AH1 float16_t + #define AH2 float16_t2 + #define AH3 float16_t3 + #define AH4 float16_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AW1 uint16_t + #define AW2 uint16_t2 + #define AW3 uint16_t3 + #define AW4 uint16_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASW1 int16_t + #define ASW2 int16_t2 + #define ASW3 int16_t3 + #define ASW4 int16_t4 + #else + #define AH1 min16float + #define AH2 min16float2 + #define AH3 min16float3 + #define AH4 min16float4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AW1 min16uint + #define AW2 min16uint2 + #define AW3 min16uint3 + #define AW4 min16uint4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASW1 min16int + #define ASW2 min16int2 + #define ASW3 min16int3 + #define ASW4 min16int4 + #endif +//============================================================================================================================== + // Need to use manual unpack to get optimal execution (don't use packed types in buffers directly). + // Unpack requires this pattern: https://gpuopen.com/first-steps-implementing-fp16/ + AH2 AH2_AU1_x(AU1 x){AF2 t=f16tof32(AU2(x&0xFFFF,x>>16));return AH2(t);} + AH4 AH4_AU2_x(AU2 x){return AH4(AH2_AU1_x(x.x),AH2_AU1_x(x.y));} + AW2 AW2_AU1_x(AU1 x){AU2 t=AU2(x&0xFFFF,x>>16);return AW2(t);} + AW4 AW4_AU2_x(AU2 x){return AW4(AW2_AU1_x(x.x),AW2_AU1_x(x.y));} + #define AH2_AU1(x) AH2_AU1_x(AU1(x)) + #define AH4_AU2(x) AH4_AU2_x(AU2(x)) + #define AW2_AU1(x) AW2_AU1_x(AU1(x)) + #define AW4_AU2(x) AW4_AU2_x(AU2(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH2_x(AH2 x){return f32tof16(x.x)+(f32tof16(x.y)<<16);} + AU2 AU2_AH4_x(AH4 x){return AU2(AU1_AH2_x(x.xy),AU1_AH2_x(x.zw));} + AU1 AU1_AW2_x(AW2 x){return AU1(x.x)+(AU1(x.y)<<16);} + AU2 AU2_AW4_x(AW4 x){return AU2(AU1_AW2_x(x.xy),AU1_AW2_x(x.zw));} + #define AU1_AH2(x) AU1_AH2_x(AH2(x)) + #define AU2_AH4(x) AU2_AH4_x(AH4(x)) + #define AU1_AW2(x) AU1_AW2_x(AW2(x)) + #define AU2_AW4(x) AU2_AW4_x(AW4(x)) +//============================================================================================================================== + #if defined(A_HLSL_6_2) && !defined(A_NO_16_BIT_CAST) + #define AW1_AH1(x) asuint16(x) + #define AW2_AH2(x) asuint16(x) + #define AW3_AH3(x) asuint16(x) + #define AW4_AH4(x) asuint16(x) + #else + #define AW1_AH1(a) AW1(f32tof16(AF1(a))) + #define AW2_AH2(a) AW2(AW1_AH1((a).x),AW1_AH1((a).y)) + #define AW3_AH3(a) AW3(AW1_AH1((a).x),AW1_AH1((a).y),AW1_AH1((a).z)) + #define AW4_AH4(a) AW4(AW1_AH1((a).x),AW1_AH1((a).y),AW1_AH1((a).z),AW1_AH1((a).w)) + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #if defined(A_HLSL_6_2) && !defined(A_NO_16_BIT_CAST) + #define AH1_AW1(x) asfloat16(x) + #define AH2_AW2(x) asfloat16(x) + #define AH3_AW3(x) asfloat16(x) + #define AH4_AW4(x) asfloat16(x) + #else + #define AH1_AW1(a) AH1(f16tof32(AU1(a))) + #define AH2_AW2(a) AH2(AH1_AW1((a).x),AH1_AW1((a).y)) + #define AH3_AW3(a) AH3(AH1_AW1((a).x),AH1_AW1((a).y),AH1_AW1((a).z)) + #define AH4_AW4(a) AH4(AH1_AW1((a).x),AH1_AW1((a).y),AH1_AW1((a).z),AH1_AW1((a).w)) + #endif +//============================================================================================================================== + AH1 AH1_x(AH1 a){return AH1(a);} + AH2 AH2_x(AH1 a){return AH2(a,a);} + AH3 AH3_x(AH1 a){return AH3(a,a,a);} + AH4 AH4_x(AH1 a){return AH4(a,a,a,a);} + #define AH1_(a) AH1_x(AH1(a)) + #define AH2_(a) AH2_x(AH1(a)) + #define AH3_(a) AH3_x(AH1(a)) + #define AH4_(a) AH4_x(AH1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AW1_x(AW1 a){return AW1(a);} + AW2 AW2_x(AW1 a){return AW2(a,a);} + AW3 AW3_x(AW1 a){return AW3(a,a,a);} + AW4 AW4_x(AW1 a){return AW4(a,a,a,a);} + #define AW1_(a) AW1_x(AW1(a)) + #define AW2_(a) AW2_x(AW1(a)) + #define AW3_(a) AW3_x(AW1(a)) + #define AW4_(a) AW4_x(AW1(a)) +//============================================================================================================================== + AW1 AAbsSW1(AW1 a){return AW1(abs(ASW1(a)));} + AW2 AAbsSW2(AW2 a){return AW2(abs(ASW2(a)));} + AW3 AAbsSW3(AW3 a){return AW3(abs(ASW3(a)));} + AW4 AAbsSW4(AW4 a){return AW4(abs(ASW4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AClampH1(AH1 x,AH1 n,AH1 m){return max(n,min(x,m));} + AH2 AClampH2(AH2 x,AH2 n,AH2 m){return max(n,min(x,m));} + AH3 AClampH3(AH3 x,AH3 n,AH3 m){return max(n,min(x,m));} + AH4 AClampH4(AH4 x,AH4 n,AH4 m){return max(n,min(x,m));} +//------------------------------------------------------------------------------------------------------------------------------ + // V_FRACT_F16 (note DX frac() is different). + AH1 AFractH1(AH1 x){return x-floor(x);} + AH2 AFractH2(AH2 x){return x-floor(x);} + AH3 AFractH3(AH3 x){return x-floor(x);} + AH4 AFractH4(AH4 x){return x-floor(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ALerpH1(AH1 x,AH1 y,AH1 a){return lerp(x,y,a);} + AH2 ALerpH2(AH2 x,AH2 y,AH2 a){return lerp(x,y,a);} + AH3 ALerpH3(AH3 x,AH3 y,AH3 a){return lerp(x,y,a);} + AH4 ALerpH4(AH4 x,AH4 y,AH4 a){return lerp(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AMax3H1(AH1 x,AH1 y,AH1 z){return max(x,max(y,z));} + AH2 AMax3H2(AH2 x,AH2 y,AH2 z){return max(x,max(y,z));} + AH3 AMax3H3(AH3 x,AH3 y,AH3 z){return max(x,max(y,z));} + AH4 AMax3H4(AH4 x,AH4 y,AH4 z){return max(x,max(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMaxSW1(AW1 a,AW1 b){return AW1(max(ASU1(a),ASU1(b)));} + AW2 AMaxSW2(AW2 a,AW2 b){return AW2(max(ASU2(a),ASU2(b)));} + AW3 AMaxSW3(AW3 a,AW3 b){return AW3(max(ASU3(a),ASU3(b)));} + AW4 AMaxSW4(AW4 a,AW4 b){return AW4(max(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AMin3H1(AH1 x,AH1 y,AH1 z){return min(x,min(y,z));} + AH2 AMin3H2(AH2 x,AH2 y,AH2 z){return min(x,min(y,z));} + AH3 AMin3H3(AH3 x,AH3 y,AH3 z){return min(x,min(y,z));} + AH4 AMin3H4(AH4 x,AH4 y,AH4 z){return min(x,min(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMinSW1(AW1 a,AW1 b){return AW1(min(ASU1(a),ASU1(b)));} + AW2 AMinSW2(AW2 a,AW2 b){return AW2(min(ASU2(a),ASU2(b)));} + AW3 AMinSW3(AW3 a,AW3 b){return AW3(min(ASU3(a),ASU3(b)));} + AW4 AMinSW4(AW4 a,AW4 b){return AW4(min(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARcpH1(AH1 x){return rcp(x);} + AH2 ARcpH2(AH2 x){return rcp(x);} + AH3 ARcpH3(AH3 x){return rcp(x);} + AH4 ARcpH4(AH4 x){return rcp(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARsqH1(AH1 x){return rsqrt(x);} + AH2 ARsqH2(AH2 x){return rsqrt(x);} + AH3 ARsqH3(AH3 x){return rsqrt(x);} + AH4 ARsqH4(AH4 x){return rsqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASatH1(AH1 x){return saturate(x);} + AH2 ASatH2(AH2 x){return saturate(x);} + AH3 ASatH3(AH3 x){return saturate(x);} + AH4 ASatH4(AH4 x){return saturate(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AShrSW1(AW1 a,AW1 b){return AW1(ASW1(a)>>ASW1(b));} + AW2 AShrSW2(AW2 a,AW2 b){return AW2(ASW2(a)>>ASW2(b));} + AW3 AShrSW3(AW3 a,AW3 b){return AW3(ASW3(a)>>ASW3(b));} + AW4 AShrSW4(AW4 a,AW4 b){return AW4(ASW4(a)>>ASW4(b));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HLSL DOUBLE +//============================================================================================================================== + #ifdef A_DUBL + #ifdef A_HLSL_6_2 + #define AD1 float64_t + #define AD2 float64_t2 + #define AD3 float64_t3 + #define AD4 float64_t4 + #else + #define AD1 double + #define AD2 double2 + #define AD3 double3 + #define AD4 double4 + #endif +//------------------------------------------------------------------------------------------------------------------------------ + AD1 AD1_x(AD1 a){return AD1(a);} + AD2 AD2_x(AD1 a){return AD2(a,a);} + AD3 AD3_x(AD1 a){return AD3(a,a,a);} + AD4 AD4_x(AD1 a){return AD4(a,a,a,a);} + #define AD1_(a) AD1_x(AD1(a)) + #define AD2_(a) AD2_x(AD1(a)) + #define AD3_(a) AD3_x(AD1(a)) + #define AD4_(a) AD4_x(AD1(a)) +//============================================================================================================================== + AD1 AFractD1(AD1 a){return a-floor(a);} + AD2 AFractD2(AD2 a){return a-floor(a);} + AD3 AFractD3(AD3 a){return a-floor(a);} + AD4 AFractD4(AD4 a){return a-floor(a);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ALerpD1(AD1 x,AD1 y,AD1 a){return lerp(x,y,a);} + AD2 ALerpD2(AD2 x,AD2 y,AD2 a){return lerp(x,y,a);} + AD3 ALerpD3(AD3 x,AD3 y,AD3 a){return lerp(x,y,a);} + AD4 ALerpD4(AD4 x,AD4 y,AD4 a){return lerp(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARcpD1(AD1 x){return rcp(x);} + AD2 ARcpD2(AD2 x){return rcp(x);} + AD3 ARcpD3(AD3 x){return rcp(x);} + AD4 ARcpD4(AD4 x){return rcp(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARsqD1(AD1 x){return rsqrt(x);} + AD2 ARsqD2(AD2 x){return rsqrt(x);} + AD3 ARsqD3(AD3 x){return rsqrt(x);} + AD4 ARsqD4(AD4 x){return rsqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ASatD1(AD1 x){return saturate(x);} + AD2 ASatD2(AD2 x){return saturate(x);} + AD3 ASatD3(AD3 x){return saturate(x);} + AD4 ASatD4(AD4 x){return saturate(x);} + #endif +//============================================================================================================================== +// HLSL WAVE +//============================================================================================================================== + #ifdef A_WAVE + // Where 'x' must be a compile time literal. + AF1 AWaveXorF1(AF1 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AF2 AWaveXorF2(AF2 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AF3 AWaveXorF3(AF3 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AF4 AWaveXorF4(AF4 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU1 AWaveXorU1(AU1 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU2 AWaveXorU1(AU2 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU3 AWaveXorU1(AU3 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU4 AWaveXorU1(AU4 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + AH2 AWaveXorH2(AH2 v,AU1 x){return AH2_AU1(WaveReadLaneAt(AU1_AH2(v),WaveGetLaneIndex()^x));} + AH4 AWaveXorH4(AH4 v,AU1 x){return AH4_AU2(WaveReadLaneAt(AU2_AH4(v),WaveGetLaneIndex()^x));} + AW2 AWaveXorW2(AW2 v,AU1 x){return AW2_AU1(WaveReadLaneAt(AU1_AW2(v),WaveGetLaneIndex()^x));} + AW4 AWaveXorW4(AW4 v,AU1 x){return AW4_AU1(WaveReadLaneAt(AU1_AW4(v),WaveGetLaneIndex()^x));} + #endif + #endif +//============================================================================================================================== +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// GPU COMMON +// +// +//============================================================================================================================== +#ifdef A_GPU + // Negative and positive infinity. + #define A_INFP_F AF1_AU1(0x7f800000u) + #define A_INFN_F AF1_AU1(0xff800000u) +//------------------------------------------------------------------------------------------------------------------------------ + // Copy sign from 's' to positive 'd'. + AF1 ACpySgnF1(AF1 d,AF1 s){return AF1_AU1(AU1_AF1(d)|(AU1_AF1(s)&AU1_(0x80000000u)));} + AF2 ACpySgnF2(AF2 d,AF2 s){return AF2_AU2(AU2_AF2(d)|(AU2_AF2(s)&AU2_(0x80000000u)));} + AF3 ACpySgnF3(AF3 d,AF3 s){return AF3_AU3(AU3_AF3(d)|(AU3_AF3(s)&AU3_(0x80000000u)));} + AF4 ACpySgnF4(AF4 d,AF4 s){return AF4_AU4(AU4_AF4(d)|(AU4_AF4(s)&AU4_(0x80000000u)));} +//------------------------------------------------------------------------------------------------------------------------------ + // Single operation to return (useful to create a mask to use in lerp for branch free logic), + // m=NaN := 0 + // m>=0 := 0 + // m<0 := 1 + // Uses the following useful floating point logic, + // saturate(+a*(-INF)==-INF) := 0 + // saturate( 0*(-INF)== NaN) := 0 + // saturate(-a*(-INF)==+INF) := 1 + AF1 ASignedF1(AF1 m){return ASatF1(m*AF1_(A_INFN_F));} + AF2 ASignedF2(AF2 m){return ASatF2(m*AF2_(A_INFN_F));} + AF3 ASignedF3(AF3 m){return ASatF3(m*AF3_(A_INFN_F));} + AF4 ASignedF4(AF4 m){return ASatF4(m*AF4_(A_INFN_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AGtZeroF1(AF1 m){return ASatF1(m*AF1_(A_INFP_F));} + AF2 AGtZeroF2(AF2 m){return ASatF2(m*AF2_(A_INFP_F));} + AF3 AGtZeroF3(AF3 m){return ASatF3(m*AF3_(A_INFP_F));} + AF4 AGtZeroF4(AF4 m){return ASatF4(m*AF4_(A_INFP_F));} +//============================================================================================================================== + #ifdef A_HALF + #ifdef A_HLSL_6_2 + #define A_INFP_H AH1_AW1((uint16_t)0x7c00u) + #define A_INFN_H AH1_AW1((uint16_t)0xfc00u) + #else + #define A_INFP_H AH1_AW1(0x7c00u) + #define A_INFN_H AH1_AW1(0xfc00u) + #endif + +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ACpySgnH1(AH1 d,AH1 s){return AH1_AW1(AW1_AH1(d)|(AW1_AH1(s)&AW1_(0x8000u)));} + AH2 ACpySgnH2(AH2 d,AH2 s){return AH2_AW2(AW2_AH2(d)|(AW2_AH2(s)&AW2_(0x8000u)));} + AH3 ACpySgnH3(AH3 d,AH3 s){return AH3_AW3(AW3_AH3(d)|(AW3_AH3(s)&AW3_(0x8000u)));} + AH4 ACpySgnH4(AH4 d,AH4 s){return AH4_AW4(AW4_AH4(d)|(AW4_AH4(s)&AW4_(0x8000u)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASignedH1(AH1 m){return ASatH1(m*AH1_(A_INFN_H));} + AH2 ASignedH2(AH2 m){return ASatH2(m*AH2_(A_INFN_H));} + AH3 ASignedH3(AH3 m){return ASatH3(m*AH3_(A_INFN_H));} + AH4 ASignedH4(AH4 m){return ASatH4(m*AH4_(A_INFN_H));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AGtZeroH1(AH1 m){return ASatH1(m*AH1_(A_INFP_H));} + AH2 AGtZeroH2(AH2 m){return ASatH2(m*AH2_(A_INFP_H));} + AH3 AGtZeroH3(AH3 m){return ASatH3(m*AH3_(A_INFP_H));} + AH4 AGtZeroH4(AH4 m){return ASatH4(m*AH4_(A_INFP_H));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [FIS] FLOAT INTEGER SORTABLE +//------------------------------------------------------------------------------------------------------------------------------ +// Float to integer sortable. +// - If sign bit=0, flip the sign bit (positives). +// - If sign bit=1, flip all bits (negatives). +// Integer sortable to float. +// - If sign bit=1, flip the sign bit (positives). +// - If sign bit=0, flip all bits (negatives). +// Has nice side effects. +// - Larger integers are more positive values. +// - Float zero is mapped to center of integers (so clear to integer zero is a nice default for atomic max usage). +// Burns 3 ops for conversion {shift,or,xor}. +//============================================================================================================================== + AU1 AFisToU1(AU1 x){return x^(( AShrSU1(x,AU1_(31)))|AU1_(0x80000000));} + AU1 AFisFromU1(AU1 x){return x^((~AShrSU1(x,AU1_(31)))|AU1_(0x80000000));} +//------------------------------------------------------------------------------------------------------------------------------ + // Just adjust high 16-bit value (useful when upper part of 32-bit word is a 16-bit float value). + AU1 AFisToHiU1(AU1 x){return x^(( AShrSU1(x,AU1_(15)))|AU1_(0x80000000));} + AU1 AFisFromHiU1(AU1 x){return x^((~AShrSU1(x,AU1_(15)))|AU1_(0x80000000));} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + AW1 AFisToW1(AW1 x){return x^(( AShrSW1(x,AW1_(15)))|AW1_(0x8000));} + AW1 AFisFromW1(AW1 x){return x^((~AShrSW1(x,AW1_(15)))|AW1_(0x8000));} +//------------------------------------------------------------------------------------------------------------------------------ + AW2 AFisToW2(AW2 x){return x^(( AShrSW2(x,AW2_(15)))|AW2_(0x8000));} + AW2 AFisFromW2(AW2 x){return x^((~AShrSW2(x,AW2_(15)))|AW2_(0x8000));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [PERM] V_PERM_B32 +//------------------------------------------------------------------------------------------------------------------------------ +// Support for V_PERM_B32 started in the 3rd generation of GCN. +//------------------------------------------------------------------------------------------------------------------------------ +// yyyyxxxx - The 'i' input. +// 76543210 +// ======== +// HGFEDCBA - Naming on permutation. +//------------------------------------------------------------------------------------------------------------------------------ +// TODO +// ==== +// - Make sure compiler optimizes this. +//============================================================================================================================== + #ifdef A_HALF + AU1 APerm0E0A(AU2 i){return((i.x )&0xffu)|((i.y<<16)&0xff0000u);} + AU1 APerm0F0B(AU2 i){return((i.x>> 8)&0xffu)|((i.y<< 8)&0xff0000u);} + AU1 APerm0G0C(AU2 i){return((i.x>>16)&0xffu)|((i.y )&0xff0000u);} + AU1 APerm0H0D(AU2 i){return((i.x>>24)&0xffu)|((i.y>> 8)&0xff0000u);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 APermHGFA(AU2 i){return((i.x )&0x000000ffu)|(i.y&0xffffff00u);} + AU1 APermHGFC(AU2 i){return((i.x>>16)&0x000000ffu)|(i.y&0xffffff00u);} + AU1 APermHGAE(AU2 i){return((i.x<< 8)&0x0000ff00u)|(i.y&0xffff00ffu);} + AU1 APermHGCE(AU2 i){return((i.x>> 8)&0x0000ff00u)|(i.y&0xffff00ffu);} + AU1 APermHAFE(AU2 i){return((i.x<<16)&0x00ff0000u)|(i.y&0xff00ffffu);} + AU1 APermHCFE(AU2 i){return((i.x )&0x00ff0000u)|(i.y&0xff00ffffu);} + AU1 APermAGFE(AU2 i){return((i.x<<24)&0xff000000u)|(i.y&0x00ffffffu);} + AU1 APermCGFE(AU2 i){return((i.x<< 8)&0xff000000u)|(i.y&0x00ffffffu);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 APermGCEA(AU2 i){return((i.x)&0x00ff00ffu)|((i.y<<8)&0xff00ff00u);} + AU1 APermGECA(AU2 i){return(((i.x)&0xffu)|((i.x>>8)&0xff00u)|((i.y<<16)&0xff0000u)|((i.y<<8)&0xff000000u));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [BUC] BYTE UNSIGNED CONVERSION +//------------------------------------------------------------------------------------------------------------------------------ +// Designed to use the optimal conversion, enables the scaling to possibly be factored into other computation. +// Works on a range of {0 to A_BUC_<32,16>}, for <32-bit, and 16-bit> respectively. +//------------------------------------------------------------------------------------------------------------------------------ +// OPCODE NOTES +// ============ +// GCN does not do UNORM or SNORM for bytes in opcodes. +// - V_CVT_F32_UBYTE{0,1,2,3} - Unsigned byte to float. +// - V_CVT_PKACC_U8_F32 - Float to unsigned byte (does bit-field insert into 32-bit integer). +// V_PERM_B32 does byte packing with ability to zero fill bytes as well. +// - Can pull out byte values from two sources, and zero fill upper 8-bits of packed hi and lo. +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABuc{0,1,2,3}{To,From}U1() - Designed for V_CVT_F32_UBYTE* and V_CVT_PKACCUM_U8_F32 ops. +// ==== ===== +// 0 : 0 +// 1 : 1 +// ... +// 255 : 255 +// : 256 (just outside the encoding range) +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABuc{0,1,2,3}{To,From}U2() - Designed for 16-bit denormal tricks and V_PERM_B32. +// ==== ===== +// 0 : 0 +// 1 : 1/512 +// 2 : 1/256 +// ... +// 64 : 1/8 +// 128 : 1/4 +// 255 : 255/512 +// : 1/2 (just outside the encoding range) +//------------------------------------------------------------------------------------------------------------------------------ +// OPTIMAL IMPLEMENTATIONS ON AMD ARCHITECTURES +// ============================================ +// r=ABuc0FromU1(i) +// V_CVT_F32_UBYTE0 r,i +// -------------------------------------------- +// r=ABuc0ToU1(d,i) +// V_CVT_PKACCUM_U8_F32 r,i,0,d +// -------------------------------------------- +// d=ABuc0FromU2(i) +// Where 'k0' is an SGPR with 0x0E0A +// Where 'k1' is an SGPR with {32768.0} packed into the lower 16-bits +// V_PERM_B32 d,i.x,i.y,k0 +// V_PK_FMA_F16 d,d,k1.x,0 +// -------------------------------------------- +// r=ABuc0ToU2(d,i) +// Where 'k0' is an SGPR with {1.0/32768.0} packed into the lower 16-bits +// Where 'k1' is an SGPR with 0x???? +// Where 'k2' is an SGPR with 0x???? +// V_PK_FMA_F16 i,i,k0.x,0 +// V_PERM_B32 r.x,i,i,k1 +// V_PERM_B32 r.y,i,i,k2 +//============================================================================================================================== + // Peak range for 32-bit and 16-bit operations. + #define A_BUC_32 (255.0) + #define A_BUC_16 (255.0/512.0) +//============================================================================================================================== + #if 1 + // Designed to be one V_CVT_PKACCUM_U8_F32. + // The extra min is required to pattern match to V_CVT_PKACCUM_U8_F32. + AU1 ABuc0ToU1(AU1 d,AF1 i){return (d&0xffffff00u)|((min(AU1(i),255u) )&(0x000000ffu));} + AU1 ABuc1ToU1(AU1 d,AF1 i){return (d&0xffff00ffu)|((min(AU1(i),255u)<< 8)&(0x0000ff00u));} + AU1 ABuc2ToU1(AU1 d,AF1 i){return (d&0xff00ffffu)|((min(AU1(i),255u)<<16)&(0x00ff0000u));} + AU1 ABuc3ToU1(AU1 d,AF1 i){return (d&0x00ffffffu)|((min(AU1(i),255u)<<24)&(0xff000000u));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed to be one V_CVT_F32_UBYTE*. + AF1 ABuc0FromU1(AU1 i){return AF1((i )&255u);} + AF1 ABuc1FromU1(AU1 i){return AF1((i>> 8)&255u);} + AF1 ABuc2FromU1(AU1 i){return AF1((i>>16)&255u);} + AF1 ABuc3FromU1(AU1 i){return AF1((i>>24)&255u);} + #endif +//============================================================================================================================== + #ifdef A_HALF + // Takes {x0,x1} and {y0,y1} and builds {{x0,y0},{x1,y1}}. + AW2 ABuc01ToW2(AH2 x,AH2 y){x*=AH2_(1.0/32768.0);y*=AH2_(1.0/32768.0); + return AW2_AU1(APermGCEA(AU2(AU1_AW2(AW2_AH2(x)),AU1_AW2(AW2_AH2(y)))));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed for 3 ops to do SOA to AOS and conversion. + AU2 ABuc0ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABuc1ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABuc2ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABuc3ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed for 2 ops to do both AOS to SOA, and conversion. + AH2 ABuc0FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)))*AH2_(32768.0);} + AH2 ABuc1FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)))*AH2_(32768.0);} + AH2 ABuc2FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)))*AH2_(32768.0);} + AH2 ABuc3FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)))*AH2_(32768.0);} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [BSC] BYTE SIGNED CONVERSION +//------------------------------------------------------------------------------------------------------------------------------ +// Similar to [BUC]. +// Works on a range of {-/+ A_BSC_<32,16>}, for <32-bit, and 16-bit> respectively. +//------------------------------------------------------------------------------------------------------------------------------ +// ENCODING (without zero-based encoding) +// ======== +// 0 = unused (can be used to mean something else) +// 1 = lowest value +// 128 = exact zero center (zero based encoding +// 255 = highest value +//------------------------------------------------------------------------------------------------------------------------------ +// Zero-based [Zb] flips the MSB bit of the byte (making 128 "exact zero" actually zero). +// This is useful if there is a desire for cleared values to decode as zero. +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABsc{0,1,2,3}{To,From}U2() - Designed for 16-bit denormal tricks and V_PERM_B32. +// ==== ===== +// 0 : -127/512 (unused) +// 1 : -126/512 +// 2 : -125/512 +// ... +// 128 : 0 +// ... +// 255 : 127/512 +// : 1/4 (just outside the encoding range) +//============================================================================================================================== + // Peak range for 32-bit and 16-bit operations. + #define A_BSC_32 (127.0) + #define A_BSC_16 (127.0/512.0) +//============================================================================================================================== + #if 1 + AU1 ABsc0ToU1(AU1 d,AF1 i){return (d&0xffffff00u)|((min(AU1(i+128.0),255u) )&(0x000000ffu));} + AU1 ABsc1ToU1(AU1 d,AF1 i){return (d&0xffff00ffu)|((min(AU1(i+128.0),255u)<< 8)&(0x0000ff00u));} + AU1 ABsc2ToU1(AU1 d,AF1 i){return (d&0xff00ffffu)|((min(AU1(i+128.0),255u)<<16)&(0x00ff0000u));} + AU1 ABsc3ToU1(AU1 d,AF1 i){return (d&0x00ffffffu)|((min(AU1(i+128.0),255u)<<24)&(0xff000000u));} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 ABsc0ToZbU1(AU1 d,AF1 i){return ((d&0xffffff00u)|((min(AU1(trunc(i)+128.0),255u) )&(0x000000ffu)))^0x00000080u;} + AU1 ABsc1ToZbU1(AU1 d,AF1 i){return ((d&0xffff00ffu)|((min(AU1(trunc(i)+128.0),255u)<< 8)&(0x0000ff00u)))^0x00008000u;} + AU1 ABsc2ToZbU1(AU1 d,AF1 i){return ((d&0xff00ffffu)|((min(AU1(trunc(i)+128.0),255u)<<16)&(0x00ff0000u)))^0x00800000u;} + AU1 ABsc3ToZbU1(AU1 d,AF1 i){return ((d&0x00ffffffu)|((min(AU1(trunc(i)+128.0),255u)<<24)&(0xff000000u)))^0x80000000u;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ABsc0FromU1(AU1 i){return AF1((i )&255u)-128.0;} + AF1 ABsc1FromU1(AU1 i){return AF1((i>> 8)&255u)-128.0;} + AF1 ABsc2FromU1(AU1 i){return AF1((i>>16)&255u)-128.0;} + AF1 ABsc3FromU1(AU1 i){return AF1((i>>24)&255u)-128.0;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ABsc0FromZbU1(AU1 i){return AF1(((i )&255u)^0x80u)-128.0;} + AF1 ABsc1FromZbU1(AU1 i){return AF1(((i>> 8)&255u)^0x80u)-128.0;} + AF1 ABsc2FromZbU1(AU1 i){return AF1(((i>>16)&255u)^0x80u)-128.0;} + AF1 ABsc3FromZbU1(AU1 i){return AF1(((i>>24)&255u)^0x80u)-128.0;} + #endif +//============================================================================================================================== + #ifdef A_HALF + // Takes {x0,x1} and {y0,y1} and builds {{x0,y0},{x1,y1}}. + AW2 ABsc01ToW2(AH2 x,AH2 y){x=x*AH2_(1.0/32768.0)+AH2_(0.25/32768.0);y=y*AH2_(1.0/32768.0)+AH2_(0.25/32768.0); + return AW2_AU1(APermGCEA(AU2(AU1_AW2(AW2_AH2(x)),AU1_AW2(AW2_AH2(y)))));} +//------------------------------------------------------------------------------------------------------------------------------ + AU2 ABsc0ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABsc1ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABsc2ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABsc3ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AU2 ABsc0ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABsc1ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABsc2ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABsc3ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH2 ABsc0FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc1FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc2FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc3FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)))*AH2_(32768.0)-AH2_(0.25);} +//------------------------------------------------------------------------------------------------------------------------------ + AH2 ABsc0FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc1FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc2FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc3FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HALF APPROXIMATIONS +//------------------------------------------------------------------------------------------------------------------------------ +// These support only positive inputs. +// Did not see value yet in specialization for range. +// Using quick testing, ended up mostly getting the same "best" approximation for various ranges. +// With hardware that can co-execute transcendentals, the value in approximations could be less than expected. +// However from a latency perspective, if execution of a transcendental is 4 clk, with no packed support, -> 8 clk total. +// And co-execution would require a compiler interleaving a lot of independent work for packed usage. +//------------------------------------------------------------------------------------------------------------------------------ +// The one Newton Raphson iteration form of rsq() was skipped (requires 6 ops total). +// Same with sqrt(), as this could be x*rsq() (7 ops). +//============================================================================================================================== + #ifdef A_HALF + // Minimize squared error across full positive range, 2 ops. + // The 0x1de2 based approximation maps {0 to 1} input maps to < 1 output. + AH1 APrxLoSqrtH1(AH1 a){return AH1_AW1((AW1_AH1(a)>>AW1_(1))+AW1_(0x1de2));} + AH2 APrxLoSqrtH2(AH2 a){return AH2_AW2((AW2_AH2(a)>>AW2_(1))+AW2_(0x1de2));} + AH3 APrxLoSqrtH3(AH3 a){return AH3_AW3((AW3_AH3(a)>>AW3_(1))+AW3_(0x1de2));} + AH4 APrxLoSqrtH4(AH4 a){return AH4_AW4((AW4_AH4(a)>>AW4_(1))+AW4_(0x1de2));} +//------------------------------------------------------------------------------------------------------------------------------ + // Lower precision estimation, 1 op. + // Minimize squared error across {smallest normal to 16384.0}. + AH1 APrxLoRcpH1(AH1 a){return AH1_AW1(AW1_(0x7784)-AW1_AH1(a));} + AH2 APrxLoRcpH2(AH2 a){return AH2_AW2(AW2_(0x7784)-AW2_AH2(a));} + AH3 APrxLoRcpH3(AH3 a){return AH3_AW3(AW3_(0x7784)-AW3_AH3(a));} + AH4 APrxLoRcpH4(AH4 a){return AH4_AW4(AW4_(0x7784)-AW4_AH4(a));} +//------------------------------------------------------------------------------------------------------------------------------ + // Medium precision estimation, one Newton Raphson iteration, 3 ops. + AH1 APrxMedRcpH1(AH1 a){AH1 b=AH1_AW1(AW1_(0x778d)-AW1_AH1(a));return b*(-b*a+AH1_(2.0));} + AH2 APrxMedRcpH2(AH2 a){AH2 b=AH2_AW2(AW2_(0x778d)-AW2_AH2(a));return b*(-b*a+AH2_(2.0));} + AH3 APrxMedRcpH3(AH3 a){AH3 b=AH3_AW3(AW3_(0x778d)-AW3_AH3(a));return b*(-b*a+AH3_(2.0));} + AH4 APrxMedRcpH4(AH4 a){AH4 b=AH4_AW4(AW4_(0x778d)-AW4_AH4(a));return b*(-b*a+AH4_(2.0));} +//------------------------------------------------------------------------------------------------------------------------------ + // Minimize squared error across {smallest normal to 16384.0}, 2 ops. + AH1 APrxLoRsqH1(AH1 a){return AH1_AW1(AW1_(0x59a3)-(AW1_AH1(a)>>AW1_(1)));} + AH2 APrxLoRsqH2(AH2 a){return AH2_AW2(AW2_(0x59a3)-(AW2_AH2(a)>>AW2_(1)));} + AH3 APrxLoRsqH3(AH3 a){return AH3_AW3(AW3_(0x59a3)-(AW3_AH3(a)>>AW3_(1)));} + AH4 APrxLoRsqH4(AH4 a){return AH4_AW4(AW4_(0x59a3)-(AW4_AH4(a)>>AW4_(1)));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// FLOAT APPROXIMATIONS +//------------------------------------------------------------------------------------------------------------------------------ +// Michal Drobot has an excellent presentation on these: "Low Level Optimizations For GCN", +// - Idea dates back to SGI, then to Quake 3, etc. +// - https://michaldrobot.files.wordpress.com/2014/05/gcn_alu_opt_digitaldragons2014.pdf +// - sqrt(x)=rsqrt(x)*x +// - rcp(x)=rsqrt(x)*rsqrt(x) for positive x +// - https://github.com/michaldrobot/ShaderFastLibs/blob/master/ShaderFastMathLib.h +//------------------------------------------------------------------------------------------------------------------------------ +// These below are from perhaps less complete searching for optimal. +// Used FP16 normal range for testing with +4096 32-bit step size for sampling error. +// So these match up well with the half approximations. +//============================================================================================================================== + AF1 APrxLoSqrtF1(AF1 a){return AF1_AU1((AU1_AF1(a)>>AU1_(1))+AU1_(0x1fbc4639));} + AF1 APrxLoRcpF1(AF1 a){return AF1_AU1(AU1_(0x7ef07ebb)-AU1_AF1(a));} + AF1 APrxMedRcpF1(AF1 a){AF1 b=AF1_AU1(AU1_(0x7ef19fff)-AU1_AF1(a));return b*(-b*a+AF1_(2.0));} + AF1 APrxLoRsqF1(AF1 a){return AF1_AU1(AU1_(0x5f347d74)-(AU1_AF1(a)>>AU1_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 APrxLoSqrtF2(AF2 a){return AF2_AU2((AU2_AF2(a)>>AU2_(1))+AU2_(0x1fbc4639));} + AF2 APrxLoRcpF2(AF2 a){return AF2_AU2(AU2_(0x7ef07ebb)-AU2_AF2(a));} + AF2 APrxMedRcpF2(AF2 a){AF2 b=AF2_AU2(AU2_(0x7ef19fff)-AU2_AF2(a));return b*(-b*a+AF2_(2.0));} + AF2 APrxLoRsqF2(AF2 a){return AF2_AU2(AU2_(0x5f347d74)-(AU2_AF2(a)>>AU2_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF3 APrxLoSqrtF3(AF3 a){return AF3_AU3((AU3_AF3(a)>>AU3_(1))+AU3_(0x1fbc4639));} + AF3 APrxLoRcpF3(AF3 a){return AF3_AU3(AU3_(0x7ef07ebb)-AU3_AF3(a));} + AF3 APrxMedRcpF3(AF3 a){AF3 b=AF3_AU3(AU3_(0x7ef19fff)-AU3_AF3(a));return b*(-b*a+AF3_(2.0));} + AF3 APrxLoRsqF3(AF3 a){return AF3_AU3(AU3_(0x5f347d74)-(AU3_AF3(a)>>AU3_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF4 APrxLoSqrtF4(AF4 a){return AF4_AU4((AU4_AF4(a)>>AU4_(1))+AU4_(0x1fbc4639));} + AF4 APrxLoRcpF4(AF4 a){return AF4_AU4(AU4_(0x7ef07ebb)-AU4_AF4(a));} + AF4 APrxMedRcpF4(AF4 a){AF4 b=AF4_AU4(AU4_(0x7ef19fff)-AU4_AF4(a));return b*(-b*a+AF4_(2.0));} + AF4 APrxLoRsqF4(AF4 a){return AF4_AU4(AU4_(0x5f347d74)-(AU4_AF4(a)>>AU4_(1)));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PQ APPROXIMATIONS +//------------------------------------------------------------------------------------------------------------------------------ +// PQ is very close to x^(1/8). The functions below Use the fast float approximation method to do +// PQ<~>Gamma2 (4th power and fast 4th root) and PQ<~>Linear (8th power and fast 8th root). Maximum error is ~0.2%. +//============================================================================================================================== +// Helpers + AF1 Quart(AF1 a) { a = a * a; return a * a;} + AF1 Oct(AF1 a) { a = a * a; a = a * a; return a * a; } + AF2 Quart(AF2 a) { a = a * a; return a * a; } + AF2 Oct(AF2 a) { a = a * a; a = a * a; return a * a; } + AF3 Quart(AF3 a) { a = a * a; return a * a; } + AF3 Oct(AF3 a) { a = a * a; a = a * a; return a * a; } + AF4 Quart(AF4 a) { a = a * a; return a * a; } + AF4 Oct(AF4 a) { a = a * a; a = a * a; return a * a; } + //------------------------------------------------------------------------------------------------------------------------------ + AF1 APrxPQToGamma2(AF1 a) { return Quart(a); } + AF1 APrxPQToLinear(AF1 a) { return Oct(a); } + AF1 APrxLoGamma2ToPQ(AF1 a) { return AF1_AU1((AU1_AF1(a) >> AU1_(2)) + AU1_(0x2F9A4E46)); } + AF1 APrxMedGamma2ToPQ(AF1 a) { AF1 b = AF1_AU1((AU1_AF1(a) >> AU1_(2)) + AU1_(0x2F9A4E46)); AF1 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF1 APrxHighGamma2ToPQ(AF1 a) { return sqrt(sqrt(a)); } + AF1 APrxLoLinearToPQ(AF1 a) { return AF1_AU1((AU1_AF1(a) >> AU1_(3)) + AU1_(0x378D8723)); } + AF1 APrxMedLinearToPQ(AF1 a) { AF1 b = AF1_AU1((AU1_AF1(a) >> AU1_(3)) + AU1_(0x378D8723)); AF1 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF1 APrxHighLinearToPQ(AF1 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF2 APrxPQToGamma2(AF2 a) { return Quart(a); } + AF2 APrxPQToLinear(AF2 a) { return Oct(a); } + AF2 APrxLoGamma2ToPQ(AF2 a) { return AF2_AU2((AU2_AF2(a) >> AU2_(2)) + AU2_(0x2F9A4E46)); } + AF2 APrxMedGamma2ToPQ(AF2 a) { AF2 b = AF2_AU2((AU2_AF2(a) >> AU2_(2)) + AU2_(0x2F9A4E46)); AF2 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF2 APrxHighGamma2ToPQ(AF2 a) { return sqrt(sqrt(a)); } + AF2 APrxLoLinearToPQ(AF2 a) { return AF2_AU2((AU2_AF2(a) >> AU2_(3)) + AU2_(0x378D8723)); } + AF2 APrxMedLinearToPQ(AF2 a) { AF2 b = AF2_AU2((AU2_AF2(a) >> AU2_(3)) + AU2_(0x378D8723)); AF2 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF2 APrxHighLinearToPQ(AF2 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF3 APrxPQToGamma2(AF3 a) { return Quart(a); } + AF3 APrxPQToLinear(AF3 a) { return Oct(a); } + AF3 APrxLoGamma2ToPQ(AF3 a) { return AF3_AU3((AU3_AF3(a) >> AU3_(2)) + AU3_(0x2F9A4E46)); } + AF3 APrxMedGamma2ToPQ(AF3 a) { AF3 b = AF3_AU3((AU3_AF3(a) >> AU3_(2)) + AU3_(0x2F9A4E46)); AF3 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF3 APrxHighGamma2ToPQ(AF3 a) { return sqrt(sqrt(a)); } + AF3 APrxLoLinearToPQ(AF3 a) { return AF3_AU3((AU3_AF3(a) >> AU3_(3)) + AU3_(0x378D8723)); } + AF3 APrxMedLinearToPQ(AF3 a) { AF3 b = AF3_AU3((AU3_AF3(a) >> AU3_(3)) + AU3_(0x378D8723)); AF3 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF3 APrxHighLinearToPQ(AF3 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF4 APrxPQToGamma2(AF4 a) { return Quart(a); } + AF4 APrxPQToLinear(AF4 a) { return Oct(a); } + AF4 APrxLoGamma2ToPQ(AF4 a) { return AF4_AU4((AU4_AF4(a) >> AU4_(2)) + AU4_(0x2F9A4E46)); } + AF4 APrxMedGamma2ToPQ(AF4 a) { AF4 b = AF4_AU4((AU4_AF4(a) >> AU4_(2)) + AU4_(0x2F9A4E46)); AF4 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF4 APrxHighGamma2ToPQ(AF4 a) { return sqrt(sqrt(a)); } + AF4 APrxLoLinearToPQ(AF4 a) { return AF4_AU4((AU4_AF4(a) >> AU4_(3)) + AU4_(0x378D8723)); } + AF4 APrxMedLinearToPQ(AF4 a) { AF4 b = AF4_AU4((AU4_AF4(a) >> AU4_(3)) + AU4_(0x378D8723)); AF4 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF4 APrxHighLinearToPQ(AF4 a) { return sqrt(sqrt(sqrt(a))); } +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PARABOLIC SIN & COS +//------------------------------------------------------------------------------------------------------------------------------ +// Approximate answers to transcendental questions. +//------------------------------------------------------------------------------------------------------------------------------ +//============================================================================================================================== + #if 1 + // Valid input range is {-1 to 1} representing {0 to 2 pi}. + // Output range is {-1/4 to 1/4} representing {-1 to 1}. + AF1 APSinF1(AF1 x){return x*abs(x)-x;} // MAD. + AF2 APSinF2(AF2 x){return x*abs(x)-x;} + AF1 APCosF1(AF1 x){x=AFractF1(x*AF1_(0.5)+AF1_(0.75));x=x*AF1_(2.0)-AF1_(1.0);return APSinF1(x);} // 3x MAD, FRACT + AF2 APCosF2(AF2 x){x=AFractF2(x*AF2_(0.5)+AF2_(0.75));x=x*AF2_(2.0)-AF2_(1.0);return APSinF2(x);} + AF2 APSinCosF1(AF1 x){AF1 y=AFractF1(x*AF1_(0.5)+AF1_(0.75));y=y*AF1_(2.0)-AF1_(1.0);return APSinF2(AF2(x,y));} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + // For a packed {sin,cos} pair, + // - Native takes 16 clocks and 4 issue slots (no packed transcendentals). + // - Parabolic takes 8 clocks and 8 issue slots (only fract is non-packed). + AH1 APSinH1(AH1 x){return x*abs(x)-x;} + AH2 APSinH2(AH2 x){return x*abs(x)-x;} // AND,FMA + AH1 APCosH1(AH1 x){x=AFractH1(x*AH1_(0.5)+AH1_(0.75));x=x*AH1_(2.0)-AH1_(1.0);return APSinH1(x);} + AH2 APCosH2(AH2 x){x=AFractH2(x*AH2_(0.5)+AH2_(0.75));x=x*AH2_(2.0)-AH2_(1.0);return APSinH2(x);} // 3x FMA, 2xFRACT, AND + AH2 APSinCosH1(AH1 x){AH1 y=AFractH1(x*AH1_(0.5)+AH1_(0.75));y=y*AH1_(2.0)-AH1_(1.0);return APSinH2(AH2(x,y));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [ZOL] ZERO ONE LOGIC +//------------------------------------------------------------------------------------------------------------------------------ +// Conditional free logic designed for easy 16-bit packing, and backwards porting to 32-bit. +//------------------------------------------------------------------------------------------------------------------------------ +// 0 := false +// 1 := true +//------------------------------------------------------------------------------------------------------------------------------ +// AndNot(x,y) -> !(x&y) .... One op. +// AndOr(x,y,z) -> (x&y)|z ... One op. +// GtZero(x) -> x>0.0 ..... One op. +// Sel(x,y,z) -> x?y:z ..... Two ops, has no precision loss. +// Signed(x) -> x<0.0 ..... One op. +// ZeroPass(x,y) -> x?0:y ..... Two ops, 'y' is a pass through safe for aliasing as integer. +//------------------------------------------------------------------------------------------------------------------------------ +// OPTIMIZATION NOTES +// ================== +// - On Vega to use 2 constants in a packed op, pass in as one AW2 or one AH2 'k.xy' and use as 'k.xx' and 'k.yy'. +// For example 'a.xy*k.xx+k.yy'. +//============================================================================================================================== + #if 1 + AU1 AZolAndU1(AU1 x,AU1 y){return min(x,y);} + AU2 AZolAndU2(AU2 x,AU2 y){return min(x,y);} + AU3 AZolAndU3(AU3 x,AU3 y){return min(x,y);} + AU4 AZolAndU4(AU4 x,AU4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AZolNotU1(AU1 x){return x^AU1_(1);} + AU2 AZolNotU2(AU2 x){return x^AU2_(1);} + AU3 AZolNotU3(AU3 x){return x^AU3_(1);} + AU4 AZolNotU4(AU4 x){return x^AU4_(1);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AZolOrU1(AU1 x,AU1 y){return max(x,y);} + AU2 AZolOrU2(AU2 x,AU2 y){return max(x,y);} + AU3 AZolOrU3(AU3 x,AU3 y){return max(x,y);} + AU4 AZolOrU4(AU4 x,AU4 y){return max(x,y);} +//============================================================================================================================== + AU1 AZolF1ToU1(AF1 x){return AU1(x);} + AU2 AZolF2ToU2(AF2 x){return AU2(x);} + AU3 AZolF3ToU3(AF3 x){return AU3(x);} + AU4 AZolF4ToU4(AF4 x){return AU4(x);} +//------------------------------------------------------------------------------------------------------------------------------ + // 2 ops, denormals don't work in 32-bit on PC (and if they are enabled, OMOD is disabled). + AU1 AZolNotF1ToU1(AF1 x){return AU1(AF1_(1.0)-x);} + AU2 AZolNotF2ToU2(AF2 x){return AU2(AF2_(1.0)-x);} + AU3 AZolNotF3ToU3(AF3 x){return AU3(AF3_(1.0)-x);} + AU4 AZolNotF4ToU4(AF4 x){return AU4(AF4_(1.0)-x);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolU1ToF1(AU1 x){return AF1(x);} + AF2 AZolU2ToF2(AU2 x){return AF2(x);} + AF3 AZolU3ToF3(AU3 x){return AF3(x);} + AF4 AZolU4ToF4(AU4 x){return AF4(x);} +//============================================================================================================================== + AF1 AZolAndF1(AF1 x,AF1 y){return min(x,y);} + AF2 AZolAndF2(AF2 x,AF2 y){return min(x,y);} + AF3 AZolAndF3(AF3 x,AF3 y){return min(x,y);} + AF4 AZolAndF4(AF4 x,AF4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ASolAndNotF1(AF1 x,AF1 y){return (-x)*y+AF1_(1.0);} + AF2 ASolAndNotF2(AF2 x,AF2 y){return (-x)*y+AF2_(1.0);} + AF3 ASolAndNotF3(AF3 x,AF3 y){return (-x)*y+AF3_(1.0);} + AF4 ASolAndNotF4(AF4 x,AF4 y){return (-x)*y+AF4_(1.0);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolAndOrF1(AF1 x,AF1 y,AF1 z){return ASatF1(x*y+z);} + AF2 AZolAndOrF2(AF2 x,AF2 y,AF2 z){return ASatF2(x*y+z);} + AF3 AZolAndOrF3(AF3 x,AF3 y,AF3 z){return ASatF3(x*y+z);} + AF4 AZolAndOrF4(AF4 x,AF4 y,AF4 z){return ASatF4(x*y+z);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolGtZeroF1(AF1 x){return ASatF1(x*AF1_(A_INFP_F));} + AF2 AZolGtZeroF2(AF2 x){return ASatF2(x*AF2_(A_INFP_F));} + AF3 AZolGtZeroF3(AF3 x){return ASatF3(x*AF3_(A_INFP_F));} + AF4 AZolGtZeroF4(AF4 x){return ASatF4(x*AF4_(A_INFP_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolNotF1(AF1 x){return AF1_(1.0)-x;} + AF2 AZolNotF2(AF2 x){return AF2_(1.0)-x;} + AF3 AZolNotF3(AF3 x){return AF3_(1.0)-x;} + AF4 AZolNotF4(AF4 x){return AF4_(1.0)-x;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolOrF1(AF1 x,AF1 y){return max(x,y);} + AF2 AZolOrF2(AF2 x,AF2 y){return max(x,y);} + AF3 AZolOrF3(AF3 x,AF3 y){return max(x,y);} + AF4 AZolOrF4(AF4 x,AF4 y){return max(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolSelF1(AF1 x,AF1 y,AF1 z){AF1 r=(-x)*z+z;return x*y+r;} + AF2 AZolSelF2(AF2 x,AF2 y,AF2 z){AF2 r=(-x)*z+z;return x*y+r;} + AF3 AZolSelF3(AF3 x,AF3 y,AF3 z){AF3 r=(-x)*z+z;return x*y+r;} + AF4 AZolSelF4(AF4 x,AF4 y,AF4 z){AF4 r=(-x)*z+z;return x*y+r;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolSignedF1(AF1 x){return ASatF1(x*AF1_(A_INFN_F));} + AF2 AZolSignedF2(AF2 x){return ASatF2(x*AF2_(A_INFN_F));} + AF3 AZolSignedF3(AF3 x){return ASatF3(x*AF3_(A_INFN_F));} + AF4 AZolSignedF4(AF4 x){return ASatF4(x*AF4_(A_INFN_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolZeroPassF1(AF1 x,AF1 y){return AF1_AU1((AU1_AF1(x)!=AU1_(0))?AU1_(0):AU1_AF1(y));} + AF2 AZolZeroPassF2(AF2 x,AF2 y){return AF2_AU2((AU2_AF2(x)!=AU2_(0))?AU2_(0):AU2_AF2(y));} + AF3 AZolZeroPassF3(AF3 x,AF3 y){return AF3_AU3((AU3_AF3(x)!=AU3_(0))?AU3_(0):AU3_AF3(y));} + AF4 AZolZeroPassF4(AF4 x,AF4 y){return AF4_AU4((AU4_AF4(x)!=AU4_(0))?AU4_(0):AU4_AF4(y));} + #endif +//============================================================================================================================== + #ifdef A_HALF + AW1 AZolAndW1(AW1 x,AW1 y){return min(x,y);} + AW2 AZolAndW2(AW2 x,AW2 y){return min(x,y);} + AW3 AZolAndW3(AW3 x,AW3 y){return min(x,y);} + AW4 AZolAndW4(AW4 x,AW4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AZolNotW1(AW1 x){return x^AW1_(1);} + AW2 AZolNotW2(AW2 x){return x^AW2_(1);} + AW3 AZolNotW3(AW3 x){return x^AW3_(1);} + AW4 AZolNotW4(AW4 x){return x^AW4_(1);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AZolOrW1(AW1 x,AW1 y){return max(x,y);} + AW2 AZolOrW2(AW2 x,AW2 y){return max(x,y);} + AW3 AZolOrW3(AW3 x,AW3 y){return max(x,y);} + AW4 AZolOrW4(AW4 x,AW4 y){return max(x,y);} +//============================================================================================================================== + // Uses denormal trick. + AW1 AZolH1ToW1(AH1 x){return AW1_AH1(x*AH1_AW1(AW1_(1)));} + AW2 AZolH2ToW2(AH2 x){return AW2_AH2(x*AH2_AW2(AW2_(1)));} + AW3 AZolH3ToW3(AH3 x){return AW3_AH3(x*AH3_AW3(AW3_(1)));} + AW4 AZolH4ToW4(AH4 x){return AW4_AH4(x*AH4_AW4(AW4_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + // AMD arch lacks a packed conversion opcode. + AH1 AZolW1ToH1(AW1 x){return AH1_AW1(x*AW1_AH1(AH1_(1.0)));} + AH2 AZolW2ToH2(AW2 x){return AH2_AW2(x*AW2_AH2(AH2_(1.0)));} + AH3 AZolW1ToH3(AW3 x){return AH3_AW3(x*AW3_AH3(AH3_(1.0)));} + AH4 AZolW2ToH4(AW4 x){return AH4_AW4(x*AW4_AH4(AH4_(1.0)));} +//============================================================================================================================== + AH1 AZolAndH1(AH1 x,AH1 y){return min(x,y);} + AH2 AZolAndH2(AH2 x,AH2 y){return min(x,y);} + AH3 AZolAndH3(AH3 x,AH3 y){return min(x,y);} + AH4 AZolAndH4(AH4 x,AH4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASolAndNotH1(AH1 x,AH1 y){return (-x)*y+AH1_(1.0);} + AH2 ASolAndNotH2(AH2 x,AH2 y){return (-x)*y+AH2_(1.0);} + AH3 ASolAndNotH3(AH3 x,AH3 y){return (-x)*y+AH3_(1.0);} + AH4 ASolAndNotH4(AH4 x,AH4 y){return (-x)*y+AH4_(1.0);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolAndOrH1(AH1 x,AH1 y,AH1 z){return ASatH1(x*y+z);} + AH2 AZolAndOrH2(AH2 x,AH2 y,AH2 z){return ASatH2(x*y+z);} + AH3 AZolAndOrH3(AH3 x,AH3 y,AH3 z){return ASatH3(x*y+z);} + AH4 AZolAndOrH4(AH4 x,AH4 y,AH4 z){return ASatH4(x*y+z);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolGtZeroH1(AH1 x){return ASatH1(x*AH1_(A_INFP_H));} + AH2 AZolGtZeroH2(AH2 x){return ASatH2(x*AH2_(A_INFP_H));} + AH3 AZolGtZeroH3(AH3 x){return ASatH3(x*AH3_(A_INFP_H));} + AH4 AZolGtZeroH4(AH4 x){return ASatH4(x*AH4_(A_INFP_H));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolNotH1(AH1 x){return AH1_(1.0)-x;} + AH2 AZolNotH2(AH2 x){return AH2_(1.0)-x;} + AH3 AZolNotH3(AH3 x){return AH3_(1.0)-x;} + AH4 AZolNotH4(AH4 x){return AH4_(1.0)-x;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolOrH1(AH1 x,AH1 y){return max(x,y);} + AH2 AZolOrH2(AH2 x,AH2 y){return max(x,y);} + AH3 AZolOrH3(AH3 x,AH3 y){return max(x,y);} + AH4 AZolOrH4(AH4 x,AH4 y){return max(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolSelH1(AH1 x,AH1 y,AH1 z){AH1 r=(-x)*z+z;return x*y+r;} + AH2 AZolSelH2(AH2 x,AH2 y,AH2 z){AH2 r=(-x)*z+z;return x*y+r;} + AH3 AZolSelH3(AH3 x,AH3 y,AH3 z){AH3 r=(-x)*z+z;return x*y+r;} + AH4 AZolSelH4(AH4 x,AH4 y,AH4 z){AH4 r=(-x)*z+z;return x*y+r;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolSignedH1(AH1 x){return ASatH1(x*AH1_(A_INFN_H));} + AH2 AZolSignedH2(AH2 x){return ASatH2(x*AH2_(A_INFN_H));} + AH3 AZolSignedH3(AH3 x){return ASatH3(x*AH3_(A_INFN_H));} + AH4 AZolSignedH4(AH4 x){return ASatH4(x*AH4_(A_INFN_H));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// COLOR CONVERSIONS +//------------------------------------------------------------------------------------------------------------------------------ +// These are all linear to/from some other space (where 'linear' has been shortened out of the function name). +// So 'ToGamma' is 'LinearToGamma', and 'FromGamma' is 'LinearFromGamma'. +// These are branch free implementations. +// The AToSrgbF1() function is useful for stores for compute shaders for GPUs without hardware linear->sRGB store conversion. +//------------------------------------------------------------------------------------------------------------------------------ +// TRANSFER FUNCTIONS +// ================== +// 709 ..... Rec709 used for some HDTVs +// Gamma ... Typically 2.2 for some PC displays, or 2.4-2.5 for CRTs, or 2.2 FreeSync2 native +// Pq ...... PQ native for HDR10 +// Srgb .... The sRGB output, typical of PC displays, useful for 10-bit output, or storing to 8-bit UNORM without SRGB type +// Two ..... Gamma 2.0, fastest conversion (useful for intermediate pass approximations) +// Three ... Gamma 3.0, less fast, but good for HDR. +//------------------------------------------------------------------------------------------------------------------------------ +// KEEPING TO SPEC +// =============== +// Both Rec.709 and sRGB have a linear segment which as spec'ed would intersect the curved segment 2 times. +// (a.) For 8-bit sRGB, steps {0 to 10.3} are in the linear region (4% of the encoding range). +// (b.) For 8-bit 709, steps {0 to 20.7} are in the linear region (8% of the encoding range). +// Also there is a slight step in the transition regions. +// Precision of the coefficients in the spec being the likely cause. +// Main usage case of the sRGB code is to do the linear->sRGB converstion in a compute shader before store. +// This is to work around lack of hardware (typically only ROP does the conversion for free). +// To "correct" the linear segment, would be to introduce error, because hardware decode of sRGB->linear is fixed (and free). +// So this header keeps with the spec. +// For linear->sRGB transforms, the linear segment in some respects reduces error, because rounding in that region is linear. +// Rounding in the curved region in hardware (and fast software code) introduces error due to rounding in non-linear. +//------------------------------------------------------------------------------------------------------------------------------ +// FOR PQ +// ====== +// Both input and output is {0.0-1.0}, and where output 1.0 represents 10000.0 cd/m^2. +// All constants are only specified to FP32 precision. +// External PQ source reference, +// - https://github.com/ampas/aces-dev/blob/master/transforms/ctl/utilities/ACESlib.Utilities_Color.a1.0.1.ctl +//------------------------------------------------------------------------------------------------------------------------------ +// PACKED VERSIONS +// =============== +// These are the A*H2() functions. +// There is no PQ functions as FP16 seemed to not have enough precision for the conversion. +// The remaining functions are "good enough" for 8-bit, and maybe 10-bit if not concerned about a few 1-bit errors. +// Precision is lowest in the 709 conversion, higher in sRGB, higher still in Two and Gamma (when using 2.2 at least). +//------------------------------------------------------------------------------------------------------------------------------ +// NOTES +// ===== +// Could be faster for PQ conversions to be in ALU or a texture lookup depending on usage case. +//============================================================================================================================== + #if 1 + AF1 ATo709F1(AF1 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AF2 ATo709F2(AF2 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AF3 ATo709F3(AF3 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + // Note 'rcpX' is '1/x', where the 'x' is what would be used in AFromGamma(). + AF1 AToGammaF1(AF1 c,AF1 rcpX){return pow(c,AF1_(rcpX));} + AF2 AToGammaF2(AF2 c,AF1 rcpX){return pow(c,AF2_(rcpX));} + AF3 AToGammaF3(AF3 c,AF1 rcpX){return pow(c,AF3_(rcpX));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToPqF1(AF1 x){AF1 p=pow(x,AF1_(0.159302)); + return pow((AF1_(0.835938)+AF1_(18.8516)*p)/(AF1_(1.0)+AF1_(18.6875)*p),AF1_(78.8438));} + AF2 AToPqF1(AF2 x){AF2 p=pow(x,AF2_(0.159302)); + return pow((AF2_(0.835938)+AF2_(18.8516)*p)/(AF2_(1.0)+AF2_(18.6875)*p),AF2_(78.8438));} + AF3 AToPqF1(AF3 x){AF3 p=pow(x,AF3_(0.159302)); + return pow((AF3_(0.835938)+AF3_(18.8516)*p)/(AF3_(1.0)+AF3_(18.6875)*p),AF3_(78.8438));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToSrgbF1(AF1 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AF2 AToSrgbF2(AF2 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AF3 AToSrgbF3(AF3 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToTwoF1(AF1 c){return sqrt(c);} + AF2 AToTwoF2(AF2 c){return sqrt(c);} + AF3 AToTwoF3(AF3 c){return sqrt(c);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToThreeF1(AF1 c){return pow(c,AF1_(1.0/3.0));} + AF2 AToThreeF2(AF2 c){return pow(c,AF2_(1.0/3.0));} + AF3 AToThreeF3(AF3 c){return pow(c,AF3_(1.0/3.0));} + #endif +//============================================================================================================================== + #if 1 + // Unfortunately median won't work here. + AF1 AFrom709F1(AF1 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF1(AZolSignedF1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AF2 AFrom709F2(AF2 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF2(AZolSignedF2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AF3 AFrom709F3(AF3 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF3(AZolSignedF3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromGammaF1(AF1 c,AF1 x){return pow(c,AF1_(x));} + AF2 AFromGammaF2(AF2 c,AF1 x){return pow(c,AF2_(x));} + AF3 AFromGammaF3(AF3 c,AF1 x){return pow(c,AF3_(x));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromPqF1(AF1 x){AF1 p=pow(x,AF1_(0.0126833)); + return pow(ASatF1(p-AF1_(0.835938))/(AF1_(18.8516)-AF1_(18.6875)*p),AF1_(6.27739));} + AF2 AFromPqF1(AF2 x){AF2 p=pow(x,AF2_(0.0126833)); + return pow(ASatF2(p-AF2_(0.835938))/(AF2_(18.8516)-AF2_(18.6875)*p),AF2_(6.27739));} + AF3 AFromPqF1(AF3 x){AF3 p=pow(x,AF3_(0.0126833)); + return pow(ASatF3(p-AF3_(0.835938))/(AF3_(18.8516)-AF3_(18.6875)*p),AF3_(6.27739));} +//------------------------------------------------------------------------------------------------------------------------------ + // Unfortunately median won't work here. + AF1 AFromSrgbF1(AF1 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF1(AZolSignedF1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AF2 AFromSrgbF2(AF2 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF2(AZolSignedF2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AF3 AFromSrgbF3(AF3 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF3(AZolSignedF3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromTwoF1(AF1 c){return c*c;} + AF2 AFromTwoF2(AF2 c){return c*c;} + AF3 AFromTwoF3(AF3 c){return c*c;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromThreeF1(AF1 c){return c*c*c;} + AF2 AFromThreeF2(AF2 c){return c*c*c;} + AF3 AFromThreeF3(AF3 c){return c*c*c;} + #endif +//============================================================================================================================== + #ifdef A_HALF + AH1 ATo709H1(AH1 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AH2 ATo709H2(AH2 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AH3 ATo709H3(AH3 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToGammaH1(AH1 c,AH1 rcpX){return pow(c,AH1_(rcpX));} + AH2 AToGammaH2(AH2 c,AH1 rcpX){return pow(c,AH2_(rcpX));} + AH3 AToGammaH3(AH3 c,AH1 rcpX){return pow(c,AH3_(rcpX));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToSrgbH1(AH1 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AH2 AToSrgbH2(AH2 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AH3 AToSrgbH3(AH3 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToTwoH1(AH1 c){return sqrt(c);} + AH2 AToTwoH2(AH2 c){return sqrt(c);} + AH3 AToTwoH3(AH3 c){return sqrt(c);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToThreeF1(AH1 c){return pow(c,AH1_(1.0/3.0));} + AH2 AToThreeF2(AH2 c){return pow(c,AH2_(1.0/3.0));} + AH3 AToThreeF3(AH3 c){return pow(c,AH3_(1.0/3.0));} + #endif +//============================================================================================================================== + #ifdef A_HALF + AH1 AFrom709H1(AH1 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH1(AZolSignedH1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AH2 AFrom709H2(AH2 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH2(AZolSignedH2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AH3 AFrom709H3(AH3 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH3(AZolSignedH3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromGammaH1(AH1 c,AH1 x){return pow(c,AH1_(x));} + AH2 AFromGammaH2(AH2 c,AH1 x){return pow(c,AH2_(x));} + AH3 AFromGammaH3(AH3 c,AH1 x){return pow(c,AH3_(x));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AHromSrgbF1(AH1 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH1(AZolSignedH1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AH2 AHromSrgbF2(AH2 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH2(AZolSignedH2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AH3 AHromSrgbF3(AH3 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH3(AZolSignedH3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromTwoH1(AH1 c){return c*c;} + AH2 AFromTwoH2(AH2 c){return c*c;} + AH3 AFromTwoH3(AH3 c){return c*c;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromThreeH1(AH1 c){return c*c*c;} + AH2 AFromThreeH2(AH2 c){return c*c*c;} + AH3 AFromThreeH3(AH3 c){return c*c*c;} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CS REMAP +//============================================================================================================================== + // Simple remap 64x1 to 8x8 with rotated 2x2 pixel quads in quad linear. + // 543210 + // ====== + // ..xxx. + // yy...y + AU2 ARmp8x8(AU1 a){return AU2(ABfe(a,1u,3u),ABfiM(ABfe(a,3u,3u),a,1u));} +//============================================================================================================================== + // More complex remap 64x1 to 8x8 which is necessary for 2D wave reductions. + // 543210 + // ====== + // .xx..x + // y..yy. + // Details, + // LANE TO 8x8 MAPPING + // =================== + // 00 01 08 09 10 11 18 19 + // 02 03 0a 0b 12 13 1a 1b + // 04 05 0c 0d 14 15 1c 1d + // 06 07 0e 0f 16 17 1e 1f + // 20 21 28 29 30 31 38 39 + // 22 23 2a 2b 32 33 3a 3b + // 24 25 2c 2d 34 35 3c 3d + // 26 27 2e 2f 36 37 3e 3f + AU2 ARmpRed8x8(AU1 a){return AU2(ABfiM(ABfe(a,2u,3u),a,1u),ABfiM(ABfe(a,3u,3u),ABfe(a,1u,2u),2u));} +//============================================================================================================================== + #ifdef A_HALF + AW2 ARmp8x8H(AU1 a){return AW2(ABfe(a,1u,3u),ABfiM(ABfe(a,3u,3u),a,1u));} + AW2 ARmpRed8x8H(AU1 a){return AW2(ABfiM(ABfe(a,2u,3u),a,1u),ABfiM(ABfe(a,3u,3u),ABfe(a,1u,2u),2u));} + #endif +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// REFERENCE +// +//------------------------------------------------------------------------------------------------------------------------------ +// IEEE FLOAT RULES +// ================ +// - saturate(NaN)=0, saturate(-INF)=0, saturate(+INF)=1 +// - {+/-}0 * {+/-}INF = NaN +// - -INF + (+INF) = NaN +// - {+/-}0 / {+/-}0 = NaN +// - {+/-}INF / {+/-}INF = NaN +// - a<(-0) := sqrt(a) = NaN (a=-0.0 won't NaN) +// - 0 == -0 +// - 4/0 = +INF +// - 4/-0 = -INF +// - 4+INF = +INF +// - 4-INF = -INF +// - 4*(+INF) = +INF +// - 4*(-INF) = -INF +// - -4*(+INF) = -INF +// - sqrt(+INF) = +INF +//------------------------------------------------------------------------------------------------------------------------------ +// FP16 ENCODING +// ============= +// fedcba9876543210 +// ---------------- +// ......mmmmmmmmmm 10-bit mantissa (encodes 11-bit 0.5 to 1.0 except for denormals) +// .eeeee.......... 5-bit exponent +// .00000.......... denormals +// .00001.......... -14 exponent +// .11110.......... 15 exponent +// .111110000000000 infinity +// .11111nnnnnnnnnn NaN with n!=0 +// s............... sign +//------------------------------------------------------------------------------------------------------------------------------ +// FP16/INT16 ALIASING DENORMAL +// ============================ +// 11-bit unsigned integers alias with half float denormal/normal values, +// 1 = 2^(-24) = 1/16777216 ....................... first denormal value +// 2 = 2^(-23) +// ... +// 1023 = 2^(-14)*(1-2^(-10)) = 2^(-14)*(1-1/1024) ... last denormal value +// 1024 = 2^(-14) = 1/16384 .......................... first normal value that still maps to integers +// 2047 .............................................. last normal value that still maps to integers +// Scaling limits, +// 2^15 = 32768 ...................................... largest power of 2 scaling +// Largest pow2 conversion mapping is at *32768, +// 1 : 2^(-9) = 1/512 +// 2 : 1/256 +// 4 : 1/128 +// 8 : 1/64 +// 16 : 1/32 +// 32 : 1/16 +// 64 : 1/8 +// 128 : 1/4 +// 256 : 1/2 +// 512 : 1 +// 1024 : 2 +// 2047 : a little less than 4 +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// GPU/CPU PORTABILITY +// +// +//------------------------------------------------------------------------------------------------------------------------------ +// This is the GPU implementation. +// See the CPU implementation for docs. +//============================================================================================================================== +#ifdef A_GPU + #define A_TRUE true + #define A_FALSE false + #define A_STATIC +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR ARGUMENT/RETURN/INITIALIZATION PORTABILITY +//============================================================================================================================== + #define retAD2 AD2 + #define retAD3 AD3 + #define retAD4 AD4 + #define retAF2 AF2 + #define retAF3 AF3 + #define retAF4 AF4 + #define retAL2 AL2 + #define retAL3 AL3 + #define retAL4 AL4 + #define retAU2 AU2 + #define retAU3 AU3 + #define retAU4 AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define inAD2 in AD2 + #define inAD3 in AD3 + #define inAD4 in AD4 + #define inAF2 in AF2 + #define inAF3 in AF3 + #define inAF4 in AF4 + #define inAL2 in AL2 + #define inAL3 in AL3 + #define inAL4 in AL4 + #define inAU2 in AU2 + #define inAU3 in AU3 + #define inAU4 in AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define inoutAD2 inout AD2 + #define inoutAD3 inout AD3 + #define inoutAD4 inout AD4 + #define inoutAF2 inout AF2 + #define inoutAF3 inout AF3 + #define inoutAF4 inout AF4 + #define inoutAL2 inout AL2 + #define inoutAL3 inout AL3 + #define inoutAL4 inout AL4 + #define inoutAU2 inout AU2 + #define inoutAU3 inout AU3 + #define inoutAU4 inout AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define outAD2 out AD2 + #define outAD3 out AD3 + #define outAD4 out AD4 + #define outAF2 out AF2 + #define outAF3 out AF3 + #define outAF4 out AF4 + #define outAL2 out AL2 + #define outAL3 out AL3 + #define outAL4 out AL4 + #define outAU2 out AU2 + #define outAU3 out AU3 + #define outAU4 out AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define varAD2(x) AD2 x + #define varAD3(x) AD3 x + #define varAD4(x) AD4 x + #define varAF2(x) AF2 x + #define varAF3(x) AF3 x + #define varAF4(x) AF4 x + #define varAL2(x) AL2 x + #define varAL3(x) AL3 x + #define varAL4(x) AL4 x + #define varAU2(x) AU2 x + #define varAU3(x) AU3 x + #define varAU4(x) AU4 x +//------------------------------------------------------------------------------------------------------------------------------ + #define initAD2(x,y) AD2(x,y) + #define initAD3(x,y,z) AD3(x,y,z) + #define initAD4(x,y,z,w) AD4(x,y,z,w) + #define initAF2(x,y) AF2(x,y) + #define initAF3(x,y,z) AF3(x,y,z) + #define initAF4(x,y,z,w) AF4(x,y,z,w) + #define initAL2(x,y) AL2(x,y) + #define initAL3(x,y,z) AL3(x,y,z) + #define initAL4(x,y,z,w) AL4(x,y,z,w) + #define initAU2(x,y) AU2(x,y) + #define initAU3(x,y,z) AU3(x,y,z) + #define initAU4(x,y,z,w) AU4(x,y,z,w) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS +//============================================================================================================================== + #define AAbsD1(a) abs(AD1(a)) + #define AAbsF1(a) abs(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ACosD1(a) cos(AD1(a)) + #define ACosF1(a) cos(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ADotD2(a,b) dot(AD2(a),AD2(b)) + #define ADotD3(a,b) dot(AD3(a),AD3(b)) + #define ADotD4(a,b) dot(AD4(a),AD4(b)) + #define ADotF2(a,b) dot(AF2(a),AF2(b)) + #define ADotF3(a,b) dot(AF3(a),AF3(b)) + #define ADotF4(a,b) dot(AF4(a),AF4(b)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AExp2D1(a) exp2(AD1(a)) + #define AExp2F1(a) exp2(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AFloorD1(a) floor(AD1(a)) + #define AFloorF1(a) floor(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ALog2D1(a) log2(AD1(a)) + #define ALog2F1(a) log2(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AMaxD1(a,b) max(a,b) + #define AMaxF1(a,b) max(a,b) + #define AMaxL1(a,b) max(a,b) + #define AMaxU1(a,b) max(a,b) +//------------------------------------------------------------------------------------------------------------------------------ + #define AMinD1(a,b) min(a,b) + #define AMinF1(a,b) min(a,b) + #define AMinL1(a,b) min(a,b) + #define AMinU1(a,b) min(a,b) +//------------------------------------------------------------------------------------------------------------------------------ + #define ASinD1(a) sin(AD1(a)) + #define ASinF1(a) sin(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ASqrtD1(a) sqrt(AD1(a)) + #define ASqrtF1(a) sqrt(AF1(a)) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS - DEPENDENT +//============================================================================================================================== + #define APowD1(a,b) pow(AD1(a),AF1(b)) + #define APowF1(a,b) pow(AF1(a),AF1(b)) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR OPS +//------------------------------------------------------------------------------------------------------------------------------ +// These are added as needed for production or prototyping, so not necessarily a complete set. +// They follow a convention of taking in a destination and also returning the destination value to increase utility. +//============================================================================================================================== + #ifdef A_DUBL + AD2 opAAbsD2(outAD2 d,inAD2 a){d=abs(a);return d;} + AD3 opAAbsD3(outAD3 d,inAD3 a){d=abs(a);return d;} + AD4 opAAbsD4(outAD4 d,inAD4 a){d=abs(a);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAAddD2(outAD2 d,inAD2 a,inAD2 b){d=a+b;return d;} + AD3 opAAddD3(outAD3 d,inAD3 a,inAD3 b){d=a+b;return d;} + AD4 opAAddD4(outAD4 d,inAD4 a,inAD4 b){d=a+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAAddOneD2(outAD2 d,inAD2 a,AD1 b){d=a+AD2_(b);return d;} + AD3 opAAddOneD3(outAD3 d,inAD3 a,AD1 b){d=a+AD3_(b);return d;} + AD4 opAAddOneD4(outAD4 d,inAD4 a,AD1 b){d=a+AD4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opACpyD2(outAD2 d,inAD2 a){d=a;return d;} + AD3 opACpyD3(outAD3 d,inAD3 a){d=a;return d;} + AD4 opACpyD4(outAD4 d,inAD4 a){d=a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opALerpD2(outAD2 d,inAD2 a,inAD2 b,inAD2 c){d=ALerpD2(a,b,c);return d;} + AD3 opALerpD3(outAD3 d,inAD3 a,inAD3 b,inAD3 c){d=ALerpD3(a,b,c);return d;} + AD4 opALerpD4(outAD4 d,inAD4 a,inAD4 b,inAD4 c){d=ALerpD4(a,b,c);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opALerpOneD2(outAD2 d,inAD2 a,inAD2 b,AD1 c){d=ALerpD2(a,b,AD2_(c));return d;} + AD3 opALerpOneD3(outAD3 d,inAD3 a,inAD3 b,AD1 c){d=ALerpD3(a,b,AD3_(c));return d;} + AD4 opALerpOneD4(outAD4 d,inAD4 a,inAD4 b,AD1 c){d=ALerpD4(a,b,AD4_(c));return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMaxD2(outAD2 d,inAD2 a,inAD2 b){d=max(a,b);return d;} + AD3 opAMaxD3(outAD3 d,inAD3 a,inAD3 b){d=max(a,b);return d;} + AD4 opAMaxD4(outAD4 d,inAD4 a,inAD4 b){d=max(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMinD2(outAD2 d,inAD2 a,inAD2 b){d=min(a,b);return d;} + AD3 opAMinD3(outAD3 d,inAD3 a,inAD3 b){d=min(a,b);return d;} + AD4 opAMinD4(outAD4 d,inAD4 a,inAD4 b){d=min(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMulD2(outAD2 d,inAD2 a,inAD2 b){d=a*b;return d;} + AD3 opAMulD3(outAD3 d,inAD3 a,inAD3 b){d=a*b;return d;} + AD4 opAMulD4(outAD4 d,inAD4 a,inAD4 b){d=a*b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMulOneD2(outAD2 d,inAD2 a,AD1 b){d=a*AD2_(b);return d;} + AD3 opAMulOneD3(outAD3 d,inAD3 a,AD1 b){d=a*AD3_(b);return d;} + AD4 opAMulOneD4(outAD4 d,inAD4 a,AD1 b){d=a*AD4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opANegD2(outAD2 d,inAD2 a){d=-a;return d;} + AD3 opANegD3(outAD3 d,inAD3 a){d=-a;return d;} + AD4 opANegD4(outAD4 d,inAD4 a){d=-a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opARcpD2(outAD2 d,inAD2 a){d=ARcpD2(a);return d;} + AD3 opARcpD3(outAD3 d,inAD3 a){d=ARcpD3(a);return d;} + AD4 opARcpD4(outAD4 d,inAD4 a){d=ARcpD4(a);return d;} + #endif +//============================================================================================================================== + AF2 opAAbsF2(outAF2 d,inAF2 a){d=abs(a);return d;} + AF3 opAAbsF3(outAF3 d,inAF3 a){d=abs(a);return d;} + AF4 opAAbsF4(outAF4 d,inAF4 a){d=abs(a);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAAddF2(outAF2 d,inAF2 a,inAF2 b){d=a+b;return d;} + AF3 opAAddF3(outAF3 d,inAF3 a,inAF3 b){d=a+b;return d;} + AF4 opAAddF4(outAF4 d,inAF4 a,inAF4 b){d=a+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAAddOneF2(outAF2 d,inAF2 a,AF1 b){d=a+AF2_(b);return d;} + AF3 opAAddOneF3(outAF3 d,inAF3 a,AF1 b){d=a+AF3_(b);return d;} + AF4 opAAddOneF4(outAF4 d,inAF4 a,AF1 b){d=a+AF4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opACpyF2(outAF2 d,inAF2 a){d=a;return d;} + AF3 opACpyF3(outAF3 d,inAF3 a){d=a;return d;} + AF4 opACpyF4(outAF4 d,inAF4 a){d=a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opALerpF2(outAF2 d,inAF2 a,inAF2 b,inAF2 c){d=ALerpF2(a,b,c);return d;} + AF3 opALerpF3(outAF3 d,inAF3 a,inAF3 b,inAF3 c){d=ALerpF3(a,b,c);return d;} + AF4 opALerpF4(outAF4 d,inAF4 a,inAF4 b,inAF4 c){d=ALerpF4(a,b,c);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opALerpOneF2(outAF2 d,inAF2 a,inAF2 b,AF1 c){d=ALerpF2(a,b,AF2_(c));return d;} + AF3 opALerpOneF3(outAF3 d,inAF3 a,inAF3 b,AF1 c){d=ALerpF3(a,b,AF3_(c));return d;} + AF4 opALerpOneF4(outAF4 d,inAF4 a,inAF4 b,AF1 c){d=ALerpF4(a,b,AF4_(c));return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMaxF2(outAF2 d,inAF2 a,inAF2 b){d=max(a,b);return d;} + AF3 opAMaxF3(outAF3 d,inAF3 a,inAF3 b){d=max(a,b);return d;} + AF4 opAMaxF4(outAF4 d,inAF4 a,inAF4 b){d=max(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMinF2(outAF2 d,inAF2 a,inAF2 b){d=min(a,b);return d;} + AF3 opAMinF3(outAF3 d,inAF3 a,inAF3 b){d=min(a,b);return d;} + AF4 opAMinF4(outAF4 d,inAF4 a,inAF4 b){d=min(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMulF2(outAF2 d,inAF2 a,inAF2 b){d=a*b;return d;} + AF3 opAMulF3(outAF3 d,inAF3 a,inAF3 b){d=a*b;return d;} + AF4 opAMulF4(outAF4 d,inAF4 a,inAF4 b){d=a*b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMulOneF2(outAF2 d,inAF2 a,AF1 b){d=a*AF2_(b);return d;} + AF3 opAMulOneF3(outAF3 d,inAF3 a,AF1 b){d=a*AF3_(b);return d;} + AF4 opAMulOneF4(outAF4 d,inAF4 a,AF1 b){d=a*AF4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opANegF2(outAF2 d,inAF2 a){d=-a;return d;} + AF3 opANegF3(outAF3 d,inAF3 a){d=-a;return d;} + AF4 opANegF4(outAF4 d,inAF4 a){d=-a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opARcpF2(outAF2 d,inAF2 a){d=ARcpF2(a);return d;} + AF3 opARcpF3(outAF3 d,inAF3 a){d=ARcpF3(a);return d;} + AF4 opARcpF4(outAF4 d,inAF4 a){d=ARcpF4(a);return d;} +#endif diff --git a/Ryujinx.Graphics.OpenGL/Effects/Shaders/ffx_fsr1.h b/Ryujinx.Graphics.OpenGL/Effects/Shaders/ffx_fsr1.h new file mode 100644 index 00000000..4e0b3d54 --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/Shaders/ffx_fsr1.h @@ -0,0 +1,1199 @@ +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// AMD FidelityFX SUPER RESOLUTION [FSR 1] ::: SPATIAL SCALING & EXTRAS - v1.20210629 +// +// +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// FidelityFX Super Resolution Sample +// +// Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved. +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files(the "Software"), to deal +// in the Software without restriction, including without limitation the rights +// to use, copy, modify, merge, publish, distribute, sublicense, and / or sell +// copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions : +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN +// THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// ABOUT +// ===== +// FSR is a collection of algorithms relating to generating a higher resolution image. +// This specific header focuses on single-image non-temporal image scaling, and related tools. +// +// The core functions are EASU and RCAS: +// [EASU] Edge Adaptive Spatial Upsampling ....... 1x to 4x area range spatial scaling, clamped adaptive elliptical filter. +// [RCAS] Robust Contrast Adaptive Sharpening .... A non-scaling variation on CAS. +// RCAS needs to be applied after EASU as a separate pass. +// +// Optional utility functions are: +// [LFGA] Linear Film Grain Applicator ........... Tool to apply film grain after scaling. +// [SRTM] Simple Reversible Tone-Mapper .......... Linear HDR {0 to FP16_MAX} to {0 to 1} and back. +// [TEPD] Temporal Energy Preserving Dither ...... Temporally energy preserving dithered {0 to 1} linear to gamma 2.0 conversion. +// See each individual sub-section for inline documentation. +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// FUNCTION PERMUTATIONS +// ===================== +// *F() ..... Single item computation with 32-bit. +// *H() ..... Single item computation with 16-bit, with packing (aka two 16-bit ops in parallel) when possible. +// *Hx2() ... Processing two items in parallel with 16-bit, easier packing. +// Not all interfaces in this file have a *Hx2() form. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [EASU] EDGE ADAPTIVE SPATIAL UPSAMPLING +// +//------------------------------------------------------------------------------------------------------------------------------ +// EASU provides a high quality spatial-only scaling at relatively low cost. +// Meaning EASU is appropiate for laptops and other low-end GPUs. +// Quality from 1x to 4x area scaling is good. +//------------------------------------------------------------------------------------------------------------------------------ +// The scalar uses a modified fast approximation to the standard lanczos(size=2) kernel. +// EASU runs in a single pass, so it applies a directionally and anisotropically adaptive radial lanczos. +// This is also kept as simple as possible to have minimum runtime. +//------------------------------------------------------------------------------------------------------------------------------ +// The lanzcos filter has negative lobes, so by itself it will introduce ringing. +// To remove all ringing, the algorithm uses the nearest 2x2 input texels as a neighborhood, +// and limits output to the minimum and maximum of that neighborhood. +//------------------------------------------------------------------------------------------------------------------------------ +// Input image requirements: +// +// Color needs to be encoded as 3 channel[red, green, blue](e.g.XYZ not supported) +// Each channel needs to be in the range[0, 1] +// Any color primaries are supported +// Display / tonemapping curve needs to be as if presenting to sRGB display or similar(e.g.Gamma 2.0) +// There should be no banding in the input +// There should be no high amplitude noise in the input +// There should be no noise in the input that is not at input pixel granularity +// For performance purposes, use 32bpp formats +//------------------------------------------------------------------------------------------------------------------------------ +// Best to apply EASU at the end of the frame after tonemapping +// but before film grain or composite of the UI. +//------------------------------------------------------------------------------------------------------------------------------ +// Example of including this header for D3D HLSL : +// +// #define A_GPU 1 +// #define A_HLSL 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of including this header for Vulkan GLSL : +// +// #define A_GPU 1 +// #define A_GLSL 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of including this header for Vulkan HLSL : +// +// #define A_GPU 1 +// #define A_HLSL 1 +// #define A_HLSL_6_2 1 +// #define A_NO_16_BIT_CAST 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of declaring the required input callbacks for GLSL : +// The callbacks need to gather4 for each color channel using the specified texture coordinate 'p'. +// EASU uses gather4 to reduce position computation logic and for free Arrays of Structures to Structures of Arrays conversion. +// +// AH4 FsrEasuRH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,0));} +// AH4 FsrEasuGH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,1));} +// AH4 FsrEasuBH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,2));} +// ... +// The FsrEasuCon function needs to be called from the CPU or GPU to set up constants. +// The difference in viewport and input image size is there to support Dynamic Resolution Scaling. +// To use FsrEasuCon() on the CPU, define A_CPU before including ffx_a and ffx_fsr1. +// Including a GPU example here, the 'con0' through 'con3' values would be stored out to a constant buffer. +// AU4 con0,con1,con2,con3; +// FsrEasuCon(con0,con1,con2,con3, +// 1920.0,1080.0, // Viewport size (top left aligned) in the input image which is to be scaled. +// 3840.0,2160.0, // The size of the input image. +// 2560.0,1440.0); // The output resolution. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CONSTANT SETUP +//============================================================================================================================== +// Call to setup required constant values (works on CPU or GPU). +A_STATIC void FsrEasuCon( +outAU4 con0, +outAU4 con1, +outAU4 con2, +outAU4 con3, +// This the rendered image resolution being upscaled +AF1 inputViewportInPixelsX, +AF1 inputViewportInPixelsY, +// This is the resolution of the resource containing the input image (useful for dynamic resolution) +AF1 inputSizeInPixelsX, +AF1 inputSizeInPixelsY, +// This is the display resolution which the input image gets upscaled to +AF1 outputSizeInPixelsX, +AF1 outputSizeInPixelsY){ + // Output integer position to a pixel position in viewport. + con0[0]=AU1_AF1(inputViewportInPixelsX*ARcpF1(outputSizeInPixelsX)); + con0[1]=AU1_AF1(inputViewportInPixelsY*ARcpF1(outputSizeInPixelsY)); + con0[2]=AU1_AF1(AF1_(0.5)*inputViewportInPixelsX*ARcpF1(outputSizeInPixelsX)-AF1_(0.5)); + con0[3]=AU1_AF1(AF1_(0.5)*inputViewportInPixelsY*ARcpF1(outputSizeInPixelsY)-AF1_(0.5)); + // Viewport pixel position to normalized image space. + // This is used to get upper-left of 'F' tap. + con1[0]=AU1_AF1(ARcpF1(inputSizeInPixelsX)); + con1[1]=AU1_AF1(ARcpF1(inputSizeInPixelsY)); + // Centers of gather4, first offset from upper-left of 'F'. + // +---+---+ + // | | | + // +--(0)--+ + // | b | c | + // +---F---+---+---+ + // | e | f | g | h | + // +--(1)--+--(2)--+ + // | i | j | k | l | + // +---+---+---+---+ + // | n | o | + // +--(3)--+ + // | | | + // +---+---+ + con1[2]=AU1_AF1(AF1_( 1.0)*ARcpF1(inputSizeInPixelsX)); + con1[3]=AU1_AF1(AF1_(-1.0)*ARcpF1(inputSizeInPixelsY)); + // These are from (0) instead of 'F'. + con2[0]=AU1_AF1(AF1_(-1.0)*ARcpF1(inputSizeInPixelsX)); + con2[1]=AU1_AF1(AF1_( 2.0)*ARcpF1(inputSizeInPixelsY)); + con2[2]=AU1_AF1(AF1_( 1.0)*ARcpF1(inputSizeInPixelsX)); + con2[3]=AU1_AF1(AF1_( 2.0)*ARcpF1(inputSizeInPixelsY)); + con3[0]=AU1_AF1(AF1_( 0.0)*ARcpF1(inputSizeInPixelsX)); + con3[1]=AU1_AF1(AF1_( 4.0)*ARcpF1(inputSizeInPixelsY)); + con3[2]=con3[3]=0;} + +//If the an offset into the input image resource +A_STATIC void FsrEasuConOffset( + outAU4 con0, + outAU4 con1, + outAU4 con2, + outAU4 con3, + // This the rendered image resolution being upscaled + AF1 inputViewportInPixelsX, + AF1 inputViewportInPixelsY, + // This is the resolution of the resource containing the input image (useful for dynamic resolution) + AF1 inputSizeInPixelsX, + AF1 inputSizeInPixelsY, + // This is the display resolution which the input image gets upscaled to + AF1 outputSizeInPixelsX, + AF1 outputSizeInPixelsY, + // This is the input image offset into the resource containing it (useful for dynamic resolution) + AF1 inputOffsetInPixelsX, + AF1 inputOffsetInPixelsY) { + FsrEasuCon(con0, con1, con2, con3, inputViewportInPixelsX, inputViewportInPixelsY, inputSizeInPixelsX, inputSizeInPixelsY, outputSizeInPixelsX, outputSizeInPixelsY); + con0[2] = AU1_AF1(AF1_(0.5) * inputViewportInPixelsX * ARcpF1(outputSizeInPixelsX) - AF1_(0.5) + inputOffsetInPixelsX); + con0[3] = AU1_AF1(AF1_(0.5) * inputViewportInPixelsY * ARcpF1(outputSizeInPixelsY) - AF1_(0.5) + inputOffsetInPixelsY); +} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 32-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(FSR_EASU_F) + // Input callback prototypes, need to be implemented by calling shader + AF4 FsrEasuRF(AF2 p); + AF4 FsrEasuGF(AF2 p); + AF4 FsrEasuBF(AF2 p); +//------------------------------------------------------------------------------------------------------------------------------ + // Filtering for a given tap for the scalar. + void FsrEasuTapF( + inout AF3 aC, // Accumulated color, with negative lobe. + inout AF1 aW, // Accumulated weight. + AF2 off, // Pixel offset from resolve position to tap. + AF2 dir, // Gradient direction. + AF2 len, // Length. + AF1 lob, // Negative lobe strength. + AF1 clp, // Clipping point. + AF3 c){ // Tap color. + // Rotate offset by direction. + AF2 v; + v.x=(off.x*( dir.x))+(off.y*dir.y); + v.y=(off.x*(-dir.y))+(off.y*dir.x); + // Anisotropy. + v*=len; + // Compute distance^2. + AF1 d2=v.x*v.x+v.y*v.y; + // Limit to the window as at corner, 2 taps can easily be outside. + d2=min(d2,clp); + // Approximation of lancos2 without sin() or rcp(), or sqrt() to get x. + // (25/16 * (2/5 * x^2 - 1)^2 - (25/16 - 1)) * (1/4 * x^2 - 1)^2 + // |_______________________________________| |_______________| + // base window + // The general form of the 'base' is, + // (a*(b*x^2-1)^2-(a-1)) + // Where 'a=1/(2*b-b^2)' and 'b' moves around the negative lobe. + AF1 wB=AF1_(2.0/5.0)*d2+AF1_(-1.0); + AF1 wA=lob*d2+AF1_(-1.0); + wB*=wB; + wA*=wA; + wB=AF1_(25.0/16.0)*wB+AF1_(-(25.0/16.0-1.0)); + AF1 w=wB*wA; + // Do weighted average. + aC+=c*w;aW+=w;} +//------------------------------------------------------------------------------------------------------------------------------ + // Accumulate direction and length. + void FsrEasuSetF( + inout AF2 dir, + inout AF1 len, + AF2 pp, + AP1 biS,AP1 biT,AP1 biU,AP1 biV, + AF1 lA,AF1 lB,AF1 lC,AF1 lD,AF1 lE){ + // Compute bilinear weight, branches factor out as predicates are compiler time immediates. + // s t + // u v + AF1 w = AF1_(0.0); + if(biS)w=(AF1_(1.0)-pp.x)*(AF1_(1.0)-pp.y); + if(biT)w= pp.x *(AF1_(1.0)-pp.y); + if(biU)w=(AF1_(1.0)-pp.x)* pp.y ; + if(biV)w= pp.x * pp.y ; + // Direction is the '+' diff. + // a + // b c d + // e + // Then takes magnitude from abs average of both sides of 'c'. + // Length converts gradient reversal to 0, smoothly to non-reversal at 1, shaped, then adding horz and vert terms. + AF1 dc=lD-lC; + AF1 cb=lC-lB; + AF1 lenX=max(abs(dc),abs(cb)); + lenX=APrxLoRcpF1(lenX); + AF1 dirX=lD-lB; + dir.x+=dirX*w; + lenX=ASatF1(abs(dirX)*lenX); + lenX*=lenX; + len+=lenX*w; + // Repeat for the y axis. + AF1 ec=lE-lC; + AF1 ca=lC-lA; + AF1 lenY=max(abs(ec),abs(ca)); + lenY=APrxLoRcpF1(lenY); + AF1 dirY=lE-lA; + dir.y+=dirY*w; + lenY=ASatF1(abs(dirY)*lenY); + lenY*=lenY; + len+=lenY*w;} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrEasuF( + out AF3 pix, + AU2 ip, // Integer pixel position in output. + AU4 con0, // Constants generated by FsrEasuCon(). + AU4 con1, + AU4 con2, + AU4 con3){ +//------------------------------------------------------------------------------------------------------------------------------ + // Get position of 'f'. + AF2 pp=AF2(ip)*AF2_AU2(con0.xy)+AF2_AU2(con0.zw); + AF2 fp=floor(pp); + pp-=fp; +//------------------------------------------------------------------------------------------------------------------------------ + // 12-tap kernel. + // b c + // e f g h + // i j k l + // n o + // Gather 4 ordering. + // a b + // r g + // For packed FP16, need either {rg} or {ab} so using the following setup for gather in all versions, + // a b <- unused (z) + // r g + // a b a b + // r g r g + // a b + // r g <- unused (z) + // Allowing dead-code removal to remove the 'z's. + AF2 p0=fp*AF2_AU2(con1.xy)+AF2_AU2(con1.zw); + // These are from p0 to avoid pulling two constants on pre-Navi hardware. + AF2 p1=p0+AF2_AU2(con2.xy); + AF2 p2=p0+AF2_AU2(con2.zw); + AF2 p3=p0+AF2_AU2(con3.xy); + AF4 bczzR=FsrEasuRF(p0); + AF4 bczzG=FsrEasuGF(p0); + AF4 bczzB=FsrEasuBF(p0); + AF4 ijfeR=FsrEasuRF(p1); + AF4 ijfeG=FsrEasuGF(p1); + AF4 ijfeB=FsrEasuBF(p1); + AF4 klhgR=FsrEasuRF(p2); + AF4 klhgG=FsrEasuGF(p2); + AF4 klhgB=FsrEasuBF(p2); + AF4 zzonR=FsrEasuRF(p3); + AF4 zzonG=FsrEasuGF(p3); + AF4 zzonB=FsrEasuBF(p3); +//------------------------------------------------------------------------------------------------------------------------------ + // Simplest multi-channel approximate luma possible (luma times 2, in 2 FMA/MAD). + AF4 bczzL=bczzB*AF4_(0.5)+(bczzR*AF4_(0.5)+bczzG); + AF4 ijfeL=ijfeB*AF4_(0.5)+(ijfeR*AF4_(0.5)+ijfeG); + AF4 klhgL=klhgB*AF4_(0.5)+(klhgR*AF4_(0.5)+klhgG); + AF4 zzonL=zzonB*AF4_(0.5)+(zzonR*AF4_(0.5)+zzonG); + // Rename. + AF1 bL=bczzL.x; + AF1 cL=bczzL.y; + AF1 iL=ijfeL.x; + AF1 jL=ijfeL.y; + AF1 fL=ijfeL.z; + AF1 eL=ijfeL.w; + AF1 kL=klhgL.x; + AF1 lL=klhgL.y; + AF1 hL=klhgL.z; + AF1 gL=klhgL.w; + AF1 oL=zzonL.z; + AF1 nL=zzonL.w; + // Accumulate for bilinear interpolation. + AF2 dir=AF2_(0.0); + AF1 len=AF1_(0.0); + FsrEasuSetF(dir,len,pp,true, false,false,false,bL,eL,fL,gL,jL); + FsrEasuSetF(dir,len,pp,false,true ,false,false,cL,fL,gL,hL,kL); + FsrEasuSetF(dir,len,pp,false,false,true ,false,fL,iL,jL,kL,nL); + FsrEasuSetF(dir,len,pp,false,false,false,true ,gL,jL,kL,lL,oL); +//------------------------------------------------------------------------------------------------------------------------------ + // Normalize with approximation, and cleanup close to zero. + AF2 dir2=dir*dir; + AF1 dirR=dir2.x+dir2.y; + AP1 zro=dirR w = -m/(n+e+w+s) +// 1 == (w*(n+e+w+s)+m)/(4*w+1) -> w = (1-m)/(n+e+w+s-4*1) +// Then chooses the 'w' which results in no clipping, limits 'w', and multiplies by the 'sharp' amount. +// This solution above has issues with MSAA input as the steps along the gradient cause edge detection issues. +// So RCAS uses 4x the maximum and 4x the minimum (depending on equation)in place of the individual taps. +// As well as switching from 'm' to either the minimum or maximum (depending on side), to help in energy conservation. +// This stabilizes RCAS. +// RCAS does a simple highpass which is normalized against the local contrast then shaped, +// 0.25 +// 0.25 -1 0.25 +// 0.25 +// This is used as a noise detection filter, to reduce the effect of RCAS on grain, and focus on real edges. +// +// GLSL example for the required callbacks : +// +// AH4 FsrRcasLoadH(ASW2 p){return AH4(imageLoad(imgSrc,ASU2(p)));} +// void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b) +// { +// //do any simple input color conversions here or leave empty if none needed +// } +// +// FsrRcasCon need to be called from the CPU or GPU to set up constants. +// Including a GPU example here, the 'con' value would be stored out to a constant buffer. +// +// AU4 con; +// FsrRcasCon(con, +// 0.0); // The scale is {0.0 := maximum sharpness, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. +// --------------- +// RCAS sharpening supports a CAS-like pass-through alpha via, +// #define FSR_RCAS_PASSTHROUGH_ALPHA 1 +// RCAS also supports a define to enable a more expensive path to avoid some sharpening of noise. +// Would suggest it is better to apply film grain after RCAS sharpening (and after scaling) instead of using this define, +// #define FSR_RCAS_DENOISE 1 +//============================================================================================================================== +// This is set at the limit of providing unnatural results for sharpening. +#define FSR_RCAS_LIMIT (0.25-(1.0/16.0)) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CONSTANT SETUP +//============================================================================================================================== +// Call to setup required constant values (works on CPU or GPU). +A_STATIC void FsrRcasCon( +outAU4 con, +// The scale is {0.0 := maximum, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. +AF1 sharpness){ + // Transform from stops to linear value. + sharpness=AExp2F1(-sharpness); + varAF2(hSharp)=initAF2(sharpness,sharpness); + con[0]=AU1_AF1(sharpness); + con[1]=AU1_AH2_AF2(hSharp); + con[2]=0; + con[3]=0;} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 32-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(FSR_RCAS_F) + // Input callback prototypes that need to be implemented by calling shader + AF4 FsrRcasLoadF(ASU2 p); + void FsrRcasInputF(inout AF1 r,inout AF1 g,inout AF1 b); +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasF( + out AF1 pixR, // Output values, non-vector so port between RcasFilter() and RcasFilterH() is easy. + out AF1 pixG, + out AF1 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AF1 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // Algorithm uses minimal 3x3 pixel neighborhood. + // b + // d e f + // h + ASU2 sp=ASU2(ip); + AF3 b=FsrRcasLoadF(sp+ASU2( 0,-1)).rgb; + AF3 d=FsrRcasLoadF(sp+ASU2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AF4 ee=FsrRcasLoadF(sp); + AF3 e=ee.rgb;pixA=ee.a; + #else + AF3 e=FsrRcasLoadF(sp).rgb; + #endif + AF3 f=FsrRcasLoadF(sp+ASU2( 1, 0)).rgb; + AF3 h=FsrRcasLoadF(sp+ASU2( 0, 1)).rgb; + // Rename (32-bit) or regroup (16-bit). + AF1 bR=b.r; + AF1 bG=b.g; + AF1 bB=b.b; + AF1 dR=d.r; + AF1 dG=d.g; + AF1 dB=d.b; + AF1 eR=e.r; + AF1 eG=e.g; + AF1 eB=e.b; + AF1 fR=f.r; + AF1 fG=f.g; + AF1 fB=f.b; + AF1 hR=h.r; + AF1 hG=h.g; + AF1 hB=h.b; + // Run optional input transform. + FsrRcasInputF(bR,bG,bB); + FsrRcasInputF(dR,dG,dB); + FsrRcasInputF(eR,eG,eB); + FsrRcasInputF(fR,fG,fB); + FsrRcasInputF(hR,hG,hB); + // Luma times 2. + AF1 bL=bB*AF1_(0.5)+(bR*AF1_(0.5)+bG); + AF1 dL=dB*AF1_(0.5)+(dR*AF1_(0.5)+dG); + AF1 eL=eB*AF1_(0.5)+(eR*AF1_(0.5)+eG); + AF1 fL=fB*AF1_(0.5)+(fR*AF1_(0.5)+fG); + AF1 hL=hB*AF1_(0.5)+(hR*AF1_(0.5)+hG); + // Noise detection. + AF1 nz=AF1_(0.25)*bL+AF1_(0.25)*dL+AF1_(0.25)*fL+AF1_(0.25)*hL-eL; + nz=ASatF1(abs(nz)*APrxMedRcpF1(AMax3F1(AMax3F1(bL,dL,eL),fL,hL)-AMin3F1(AMin3F1(bL,dL,eL),fL,hL))); + nz=AF1_(-0.5)*nz+AF1_(1.0); + // Min and max of ring. + AF1 mn4R=min(AMin3F1(bR,dR,fR),hR); + AF1 mn4G=min(AMin3F1(bG,dG,fG),hG); + AF1 mn4B=min(AMin3F1(bB,dB,fB),hB); + AF1 mx4R=max(AMax3F1(bR,dR,fR),hR); + AF1 mx4G=max(AMax3F1(bG,dG,fG),hG); + AF1 mx4B=max(AMax3F1(bB,dB,fB),hB); + // Immediate constants for peak range. + AF2 peakC=AF2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AF1 hitMinR=min(mn4R,eR)*ARcpF1(AF1_(4.0)*mx4R); + AF1 hitMinG=min(mn4G,eG)*ARcpF1(AF1_(4.0)*mx4G); + AF1 hitMinB=min(mn4B,eB)*ARcpF1(AF1_(4.0)*mx4B); + AF1 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpF1(AF1_(4.0)*mn4R+peakC.y); + AF1 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpF1(AF1_(4.0)*mn4G+peakC.y); + AF1 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpF1(AF1_(4.0)*mn4B+peakC.y); + AF1 lobeR=max(-hitMinR,hitMaxR); + AF1 lobeG=max(-hitMinG,hitMaxG); + AF1 lobeB=max(-hitMinB,hitMaxB); + AF1 lobe=max(AF1_(-FSR_RCAS_LIMIT),min(AMax3F1(lobeR,lobeG,lobeB),AF1_(0.0)))*AF1_AU1(con.x); + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AF1 rcpL=APrxMedRcpF1(AF1_(4.0)*lobe+AF1_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL; + return;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 16-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_RCAS_H) + // Input callback prototypes that need to be implemented by calling shader + AH4 FsrRcasLoadH(ASW2 p); + void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b); +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasH( + out AH1 pixR, // Output values, non-vector so port between RcasFilter() and RcasFilterH() is easy. + out AH1 pixG, + out AH1 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AH1 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // Sharpening algorithm uses minimal 3x3 pixel neighborhood. + // b + // d e f + // h + ASW2 sp=ASW2(ip); + AH3 b=FsrRcasLoadH(sp+ASW2( 0,-1)).rgb; + AH3 d=FsrRcasLoadH(sp+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee=FsrRcasLoadH(sp); + AH3 e=ee.rgb;pixA=ee.a; + #else + AH3 e=FsrRcasLoadH(sp).rgb; + #endif + AH3 f=FsrRcasLoadH(sp+ASW2( 1, 0)).rgb; + AH3 h=FsrRcasLoadH(sp+ASW2( 0, 1)).rgb; + // Rename (32-bit) or regroup (16-bit). + AH1 bR=b.r; + AH1 bG=b.g; + AH1 bB=b.b; + AH1 dR=d.r; + AH1 dG=d.g; + AH1 dB=d.b; + AH1 eR=e.r; + AH1 eG=e.g; + AH1 eB=e.b; + AH1 fR=f.r; + AH1 fG=f.g; + AH1 fB=f.b; + AH1 hR=h.r; + AH1 hG=h.g; + AH1 hB=h.b; + // Run optional input transform. + FsrRcasInputH(bR,bG,bB); + FsrRcasInputH(dR,dG,dB); + FsrRcasInputH(eR,eG,eB); + FsrRcasInputH(fR,fG,fB); + FsrRcasInputH(hR,hG,hB); + // Luma times 2. + AH1 bL=bB*AH1_(0.5)+(bR*AH1_(0.5)+bG); + AH1 dL=dB*AH1_(0.5)+(dR*AH1_(0.5)+dG); + AH1 eL=eB*AH1_(0.5)+(eR*AH1_(0.5)+eG); + AH1 fL=fB*AH1_(0.5)+(fR*AH1_(0.5)+fG); + AH1 hL=hB*AH1_(0.5)+(hR*AH1_(0.5)+hG); + // Noise detection. + AH1 nz=AH1_(0.25)*bL+AH1_(0.25)*dL+AH1_(0.25)*fL+AH1_(0.25)*hL-eL; + nz=ASatH1(abs(nz)*APrxMedRcpH1(AMax3H1(AMax3H1(bL,dL,eL),fL,hL)-AMin3H1(AMin3H1(bL,dL,eL),fL,hL))); + nz=AH1_(-0.5)*nz+AH1_(1.0); + // Min and max of ring. + AH1 mn4R=min(AMin3H1(bR,dR,fR),hR); + AH1 mn4G=min(AMin3H1(bG,dG,fG),hG); + AH1 mn4B=min(AMin3H1(bB,dB,fB),hB); + AH1 mx4R=max(AMax3H1(bR,dR,fR),hR); + AH1 mx4G=max(AMax3H1(bG,dG,fG),hG); + AH1 mx4B=max(AMax3H1(bB,dB,fB),hB); + // Immediate constants for peak range. + AH2 peakC=AH2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AH1 hitMinR=min(mn4R,eR)*ARcpH1(AH1_(4.0)*mx4R); + AH1 hitMinG=min(mn4G,eG)*ARcpH1(AH1_(4.0)*mx4G); + AH1 hitMinB=min(mn4B,eB)*ARcpH1(AH1_(4.0)*mx4B); + AH1 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpH1(AH1_(4.0)*mn4R+peakC.y); + AH1 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpH1(AH1_(4.0)*mn4G+peakC.y); + AH1 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpH1(AH1_(4.0)*mn4B+peakC.y); + AH1 lobeR=max(-hitMinR,hitMaxR); + AH1 lobeG=max(-hitMinG,hitMaxG); + AH1 lobeB=max(-hitMinB,hitMaxB); + AH1 lobe=max(AH1_(-FSR_RCAS_LIMIT),min(AMax3H1(lobeR,lobeG,lobeB),AH1_(0.0)))*AH2_AU1(con.y).x; + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AH1 rcpL=APrxMedRcpH1(AH1_(4.0)*lobe+AH1_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PACKED 16-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_RCAS_HX2) + // Input callback prototypes that need to be implemented by the calling shader + AH4 FsrRcasLoadHx2(ASW2 p); + void FsrRcasInputHx2(inout AH2 r,inout AH2 g,inout AH2 b); +//------------------------------------------------------------------------------------------------------------------------------ + // Can be used to convert from packed Structures of Arrays to Arrays of Structures for store. + void FsrRcasDepackHx2(out AH4 pix0,out AH4 pix1,AH2 pixR,AH2 pixG,AH2 pixB){ + #ifdef A_HLSL + // Invoke a slower path for DX only, since it won't allow uninitialized values. + pix0.a=pix1.a=0.0; + #endif + pix0.rgb=AH3(pixR.x,pixG.x,pixB.x); + pix1.rgb=AH3(pixR.y,pixG.y,pixB.y);} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasHx2( + // Output values are for 2 8x8 tiles in a 16x8 region. + // pix.x = left 8x8 tile + // pix.y = right 8x8 tile + // This enables later processing to easily be packed as well. + out AH2 pixR, + out AH2 pixG, + out AH2 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AH2 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // No scaling algorithm uses minimal 3x3 pixel neighborhood. + ASW2 sp0=ASW2(ip); + AH3 b0=FsrRcasLoadHx2(sp0+ASW2( 0,-1)).rgb; + AH3 d0=FsrRcasLoadHx2(sp0+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee0=FsrRcasLoadHx2(sp0); + AH3 e0=ee0.rgb;pixA.r=ee0.a; + #else + AH3 e0=FsrRcasLoadHx2(sp0).rgb; + #endif + AH3 f0=FsrRcasLoadHx2(sp0+ASW2( 1, 0)).rgb; + AH3 h0=FsrRcasLoadHx2(sp0+ASW2( 0, 1)).rgb; + ASW2 sp1=sp0+ASW2(8,0); + AH3 b1=FsrRcasLoadHx2(sp1+ASW2( 0,-1)).rgb; + AH3 d1=FsrRcasLoadHx2(sp1+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee1=FsrRcasLoadHx2(sp1); + AH3 e1=ee1.rgb;pixA.g=ee1.a; + #else + AH3 e1=FsrRcasLoadHx2(sp1).rgb; + #endif + AH3 f1=FsrRcasLoadHx2(sp1+ASW2( 1, 0)).rgb; + AH3 h1=FsrRcasLoadHx2(sp1+ASW2( 0, 1)).rgb; + // Arrays of Structures to Structures of Arrays conversion. + AH2 bR=AH2(b0.r,b1.r); + AH2 bG=AH2(b0.g,b1.g); + AH2 bB=AH2(b0.b,b1.b); + AH2 dR=AH2(d0.r,d1.r); + AH2 dG=AH2(d0.g,d1.g); + AH2 dB=AH2(d0.b,d1.b); + AH2 eR=AH2(e0.r,e1.r); + AH2 eG=AH2(e0.g,e1.g); + AH2 eB=AH2(e0.b,e1.b); + AH2 fR=AH2(f0.r,f1.r); + AH2 fG=AH2(f0.g,f1.g); + AH2 fB=AH2(f0.b,f1.b); + AH2 hR=AH2(h0.r,h1.r); + AH2 hG=AH2(h0.g,h1.g); + AH2 hB=AH2(h0.b,h1.b); + // Run optional input transform. + FsrRcasInputHx2(bR,bG,bB); + FsrRcasInputHx2(dR,dG,dB); + FsrRcasInputHx2(eR,eG,eB); + FsrRcasInputHx2(fR,fG,fB); + FsrRcasInputHx2(hR,hG,hB); + // Luma times 2. + AH2 bL=bB*AH2_(0.5)+(bR*AH2_(0.5)+bG); + AH2 dL=dB*AH2_(0.5)+(dR*AH2_(0.5)+dG); + AH2 eL=eB*AH2_(0.5)+(eR*AH2_(0.5)+eG); + AH2 fL=fB*AH2_(0.5)+(fR*AH2_(0.5)+fG); + AH2 hL=hB*AH2_(0.5)+(hR*AH2_(0.5)+hG); + // Noise detection. + AH2 nz=AH2_(0.25)*bL+AH2_(0.25)*dL+AH2_(0.25)*fL+AH2_(0.25)*hL-eL; + nz=ASatH2(abs(nz)*APrxMedRcpH2(AMax3H2(AMax3H2(bL,dL,eL),fL,hL)-AMin3H2(AMin3H2(bL,dL,eL),fL,hL))); + nz=AH2_(-0.5)*nz+AH2_(1.0); + // Min and max of ring. + AH2 mn4R=min(AMin3H2(bR,dR,fR),hR); + AH2 mn4G=min(AMin3H2(bG,dG,fG),hG); + AH2 mn4B=min(AMin3H2(bB,dB,fB),hB); + AH2 mx4R=max(AMax3H2(bR,dR,fR),hR); + AH2 mx4G=max(AMax3H2(bG,dG,fG),hG); + AH2 mx4B=max(AMax3H2(bB,dB,fB),hB); + // Immediate constants for peak range. + AH2 peakC=AH2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AH2 hitMinR=min(mn4R,eR)*ARcpH2(AH2_(4.0)*mx4R); + AH2 hitMinG=min(mn4G,eG)*ARcpH2(AH2_(4.0)*mx4G); + AH2 hitMinB=min(mn4B,eB)*ARcpH2(AH2_(4.0)*mx4B); + AH2 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpH2(AH2_(4.0)*mn4R+peakC.y); + AH2 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpH2(AH2_(4.0)*mn4G+peakC.y); + AH2 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpH2(AH2_(4.0)*mn4B+peakC.y); + AH2 lobeR=max(-hitMinR,hitMaxR); + AH2 lobeG=max(-hitMinG,hitMaxG); + AH2 lobeB=max(-hitMinB,hitMaxB); + AH2 lobe=max(AH2_(-FSR_RCAS_LIMIT),min(AMax3H2(lobeR,lobeG,lobeB),AH2_(0.0)))*AH2_(AH2_AU1(con.y).x); + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AH2 rcpL=APrxMedRcpH2(AH2_(4.0)*lobe+AH2_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [LFGA] LINEAR FILM GRAIN APPLICATOR +// +//------------------------------------------------------------------------------------------------------------------------------ +// Adding output-resolution film grain after scaling is a good way to mask both rendering and scaling artifacts. +// Suggest using tiled blue noise as film grain input, with peak noise frequency set for a specific look and feel. +// The 'Lfga*()' functions provide a convenient way to introduce grain. +// These functions limit grain based on distance to signal limits. +// This is done so that the grain is temporally energy preserving, and thus won't modify image tonality. +// Grain application should be done in a linear colorspace. +// The grain should be temporally changing, but have a temporal sum per pixel that adds to zero (non-biased). +//------------------------------------------------------------------------------------------------------------------------------ +// Usage, +// FsrLfga*( +// color, // In/out linear colorspace color {0 to 1} ranged. +// grain, // Per pixel grain texture value {-0.5 to 0.5} ranged, input is 3-channel to support colored grain. +// amount); // Amount of grain (0 to 1} ranged. +//------------------------------------------------------------------------------------------------------------------------------ +// Example if grain texture is monochrome: 'FsrLfgaF(color,AF3_(grain),amount)' +//============================================================================================================================== +#if defined(A_GPU) + // Maximum grain is the minimum distance to the signal limit. + void FsrLfgaF(inout AF3 c,AF3 t,AF1 a){c+=(t*AF3_(a))*min(AF3_(1.0)-c,c);} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + // Half precision version (slower). + void FsrLfgaH(inout AH3 c,AH3 t,AH1 a){c+=(t*AH3_(a))*min(AH3_(1.0)-c,c);} +//------------------------------------------------------------------------------------------------------------------------------ + // Packed half precision version (faster). + void FsrLfgaHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 tR,AH2 tG,AH2 tB,AH1 a){ + cR+=(tR*AH2_(a))*min(AH2_(1.0)-cR,cR);cG+=(tG*AH2_(a))*min(AH2_(1.0)-cG,cG);cB+=(tB*AH2_(a))*min(AH2_(1.0)-cB,cB);} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [SRTM] SIMPLE REVERSIBLE TONE-MAPPER +// +//------------------------------------------------------------------------------------------------------------------------------ +// This provides a way to take linear HDR color {0 to FP16_MAX} and convert it into a temporary {0 to 1} ranged post-tonemapped linear. +// The tonemapper preserves RGB ratio, which helps maintain HDR color bleed during filtering. +//------------------------------------------------------------------------------------------------------------------------------ +// Reversible tonemapper usage, +// FsrSrtm*(color); // {0 to FP16_MAX} converted to {0 to 1}. +// FsrSrtmInv*(color); // {0 to 1} converted into {0 to 32768, output peak safe for FP16}. +//============================================================================================================================== +#if defined(A_GPU) + void FsrSrtmF(inout AF3 c){c*=AF3_(ARcpF1(AMax3F1(c.r,c.g,c.b)+AF1_(1.0)));} + // The extra max solves the c=1.0 case (which is a /0). + void FsrSrtmInvF(inout AF3 c){c*=AF3_(ARcpF1(max(AF1_(1.0/32768.0),AF1_(1.0)-AMax3F1(c.r,c.g,c.b))));} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + void FsrSrtmH(inout AH3 c){c*=AH3_(ARcpH1(AMax3H1(c.r,c.g,c.b)+AH1_(1.0)));} + void FsrSrtmInvH(inout AH3 c){c*=AH3_(ARcpH1(max(AH1_(1.0/32768.0),AH1_(1.0)-AMax3H1(c.r,c.g,c.b))));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrSrtmHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB){ + AH2 rcp=ARcpH2(AMax3H2(cR,cG,cB)+AH2_(1.0));cR*=rcp;cG*=rcp;cB*=rcp;} + void FsrSrtmInvHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB){ + AH2 rcp=ARcpH2(max(AH2_(1.0/32768.0),AH2_(1.0)-AMax3H2(cR,cG,cB)));cR*=rcp;cG*=rcp;cB*=rcp;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [TEPD] TEMPORAL ENERGY PRESERVING DITHER +// +//------------------------------------------------------------------------------------------------------------------------------ +// Temporally energy preserving dithered {0 to 1} linear to gamma 2.0 conversion. +// Gamma 2.0 is used so that the conversion back to linear is just to square the color. +// The conversion comes in 8-bit and 10-bit modes, designed for output to 8-bit UNORM or 10:10:10:2 respectively. +// Given good non-biased temporal blue noise as dither input, +// the output dither will temporally conserve energy. +// This is done by choosing the linear nearest step point instead of perceptual nearest. +// See code below for details. +//------------------------------------------------------------------------------------------------------------------------------ +// DX SPEC RULES FOR FLOAT->UNORM 8-BIT CONVERSION +// =============================================== +// - Output is 'uint(floor(saturate(n)*255.0+0.5))'. +// - Thus rounding is to nearest. +// - NaN gets converted to zero. +// - INF is clamped to {0.0 to 1.0}. +//============================================================================================================================== +#if defined(A_GPU) + // Hand tuned integer position to dither value, with more values than simple checkerboard. + // Only 32-bit has enough precision for this compddation. + // Output is {0 to <1}. + AF1 FsrTepdDitF(AU2 p,AU1 f){ + AF1 x=AF1_(p.x+f); + AF1 y=AF1_(p.y); + // The 1.61803 golden ratio. + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + // Number designed to provide a good visual pattern. + AF1 b=AF1_(1.0/3.69); + x=x*a+(y*b); + return AFractF1(x);} +//------------------------------------------------------------------------------------------------------------------------------ + // This version is 8-bit gamma 2.0. + // The 'c' input is {0 to 1}. + // Output is {0 to 1} ready for image store. + void FsrTepdC8F(inout AF3 c,AF1 dit){ + AF3 n=sqrt(c); + n=floor(n*AF3_(255.0))*AF3_(1.0/255.0); + AF3 a=n*n; + AF3 b=n+AF3_(1.0/255.0);b=b*b; + // Ratio of 'a' to 'b' required to produce 'c'. + // APrxLoRcpF1() won't work here (at least for very high dynamic ranges). + // APrxMedRcpF1() is an IADD,FMA,MUL. + AF3 r=(c-b)*APrxMedRcpF3(a-b); + // Use the ratio as a cutoff to choose 'a' or 'b'. + // AGtZeroF1() is a MUL. + c=ASatF3(n+AGtZeroF3(AF3_(dit)-r)*AF3_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + // This version is 10-bit gamma 2.0. + // The 'c' input is {0 to 1}. + // Output is {0 to 1} ready for image store. + void FsrTepdC10F(inout AF3 c,AF1 dit){ + AF3 n=sqrt(c); + n=floor(n*AF3_(1023.0))*AF3_(1.0/1023.0); + AF3 a=n*n; + AF3 b=n+AF3_(1.0/1023.0);b=b*b; + AF3 r=(c-b)*APrxMedRcpF3(a-b); + c=ASatF3(n+AGtZeroF3(AF3_(dit)-r)*AF3_(1.0/1023.0));} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + AH1 FsrTepdDitH(AU2 p,AU1 f){ + AF1 x=AF1_(p.x+f); + AF1 y=AF1_(p.y); + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + AF1 b=AF1_(1.0/3.69); + x=x*a+(y*b); + return AH1(AFractF1(x));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC8H(inout AH3 c,AH1 dit){ + AH3 n=sqrt(c); + n=floor(n*AH3_(255.0))*AH3_(1.0/255.0); + AH3 a=n*n; + AH3 b=n+AH3_(1.0/255.0);b=b*b; + AH3 r=(c-b)*APrxMedRcpH3(a-b); + c=ASatH3(n+AGtZeroH3(AH3_(dit)-r)*AH3_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC10H(inout AH3 c,AH1 dit){ + AH3 n=sqrt(c); + n=floor(n*AH3_(1023.0))*AH3_(1.0/1023.0); + AH3 a=n*n; + AH3 b=n+AH3_(1.0/1023.0);b=b*b; + AH3 r=(c-b)*APrxMedRcpH3(a-b); + c=ASatH3(n+AGtZeroH3(AH3_(dit)-r)*AH3_(1.0/1023.0));} +//============================================================================================================================== + // This computes dither for positions 'p' and 'p+{8,0}'. + AH2 FsrTepdDitHx2(AU2 p,AU1 f){ + AF2 x; + x.x=AF1_(p.x+f); + x.y=x.x+AF1_(8.0); + AF1 y=AF1_(p.y); + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + AF1 b=AF1_(1.0/3.69); + x=x*AF2_(a)+AF2_(y*b); + return AH2(AFractF2(x));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC8Hx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 dit){ + AH2 nR=sqrt(cR); + AH2 nG=sqrt(cG); + AH2 nB=sqrt(cB); + nR=floor(nR*AH2_(255.0))*AH2_(1.0/255.0); + nG=floor(nG*AH2_(255.0))*AH2_(1.0/255.0); + nB=floor(nB*AH2_(255.0))*AH2_(1.0/255.0); + AH2 aR=nR*nR; + AH2 aG=nG*nG; + AH2 aB=nB*nB; + AH2 bR=nR+AH2_(1.0/255.0);bR=bR*bR; + AH2 bG=nG+AH2_(1.0/255.0);bG=bG*bG; + AH2 bB=nB+AH2_(1.0/255.0);bB=bB*bB; + AH2 rR=(cR-bR)*APrxMedRcpH2(aR-bR); + AH2 rG=(cG-bG)*APrxMedRcpH2(aG-bG); + AH2 rB=(cB-bB)*APrxMedRcpH2(aB-bB); + cR=ASatH2(nR+AGtZeroH2(dit-rR)*AH2_(1.0/255.0)); + cG=ASatH2(nG+AGtZeroH2(dit-rG)*AH2_(1.0/255.0)); + cB=ASatH2(nB+AGtZeroH2(dit-rB)*AH2_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC10Hx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 dit){ + AH2 nR=sqrt(cR); + AH2 nG=sqrt(cG); + AH2 nB=sqrt(cB); + nR=floor(nR*AH2_(1023.0))*AH2_(1.0/1023.0); + nG=floor(nG*AH2_(1023.0))*AH2_(1.0/1023.0); + nB=floor(nB*AH2_(1023.0))*AH2_(1.0/1023.0); + AH2 aR=nR*nR; + AH2 aG=nG*nG; + AH2 aB=nB*nB; + AH2 bR=nR+AH2_(1.0/1023.0);bR=bR*bR; + AH2 bG=nG+AH2_(1.0/1023.0);bG=bG*bG; + AH2 bB=nB+AH2_(1.0/1023.0);bB=bB*bB; + AH2 rR=(cR-bR)*APrxMedRcpH2(aR-bR); + AH2 rG=(cG-bG)*APrxMedRcpH2(aG-bG); + AH2 rB=(cB-bB)*APrxMedRcpH2(aB-bB); + cR=ASatH2(nR+AGtZeroH2(dit-rR)*AH2_(1.0/1023.0)); + cG=ASatH2(nG+AGtZeroH2(dit-rG)*AH2_(1.0/1023.0)); + cB=ASatH2(nB+AGtZeroH2(dit-rB)*AH2_(1.0/1023.0));} +#endif diff --git a/Ryujinx.Graphics.OpenGL/Effects/Shaders/fsr_scaling.glsl b/Ryujinx.Graphics.OpenGL/Effects/Shaders/fsr_scaling.glsl new file mode 100644 index 00000000..8e8755db --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/Shaders/fsr_scaling.glsl @@ -0,0 +1,88 @@ +#version 430 core +precision mediump float; +layout (local_size_x = 64) in; +layout(rgba8, binding = 0, location=0) uniform image2D imgOutput; +layout( location=1 ) uniform sampler2D Source; +layout( location=2 ) uniform float srcX0; +layout( location=3 ) uniform float srcX1; +layout( location=4 ) uniform float srcY0; +layout( location=5 ) uniform float srcY1; +layout( location=6 ) uniform float dstX0; +layout( location=7 ) uniform float dstX1; +layout( location=8 ) uniform float dstY0; +layout( location=9 ) uniform float dstY1; +layout( location=10 ) uniform float scaleX; +layout( location=11 ) uniform float scaleY; + +#define A_GPU 1 +#define A_GLSL 1 +#include "ffx_a.h" + +#define FSR_EASU_F 1 +AU4 con0, con1, con2, con3; +float srcW, srcH, dstW, dstH; +vec2 bLeft, tRight; + +AF2 translate(AF2 pos) { + return AF2(pos.x * scaleX, pos.y * scaleY); +} + +void setBounds(vec2 bottomLeft, vec2 topRight) { + bLeft = bottomLeft; + tRight = topRight; +} + +AF2 translateDest(AF2 pos) { + AF2 translatedPos = AF2(pos.x, pos.y); + translatedPos.x = dstX1 < dstX0 ? dstX1 - translatedPos.x : translatedPos.x; + translatedPos.y = dstY0 > dstY1 ? dstY0 + dstY1 - translatedPos.y - 1: translatedPos.y; + return translatedPos; +} + +AF4 FsrEasuRF(AF2 p) { AF4 res = textureGather(Source, translate(p), 0); return res; } +AF4 FsrEasuGF(AF2 p) { AF4 res = textureGather(Source, translate(p), 1); return res; } +AF4 FsrEasuBF(AF2 p) { AF4 res = textureGather(Source, translate(p), 2); return res; } + +#include "ffx_fsr1.h" + +float insideBox(vec2 v) { + vec2 s = step(bLeft, v) - step(tRight, v); + return s.x * s.y; +} + +void CurrFilter(AU2 pos) +{ + if((insideBox(vec2(pos.x, pos.y))) == 0) { + imageStore(imgOutput, ASU2(pos.x, pos.y), AF4(0,0,0,1)); + return; + } + AF3 c; + FsrEasuF(c, AU2(pos.x - bLeft.x, pos.y - bLeft.y), con0, con1, con2, con3); + imageStore(imgOutput, ASU2(translateDest(pos)), AF4(c, 1)); +} + +void main() { + srcW = abs(srcX1 - srcX0); + srcH = abs(srcY1 - srcY0); + dstW = abs(dstX1 - dstX0); + dstH = abs(dstY1 - dstY0); + + AU2 gxy = ARmp8x8(gl_LocalInvocationID.x) + AU2(gl_WorkGroupID.x << 4u, gl_WorkGroupID.y << 4u); + + setBounds(vec2(dstX0 < dstX1 ? dstX0 : dstX1, dstY0 < dstY1 ? dstY0 : dstY1), + vec2(dstX1 > dstX0 ? dstX1 : dstX0, dstY1 > dstY0 ? dstY1 : dstY0)); + + // Upscaling + FsrEasuCon(con0, con1, con2, con3, + srcW, srcH, // Viewport size (top left aligned) in the input image which is to be scaled. + srcW, srcH, // The size of the input image. + dstW, dstH); // The output resolution. + + CurrFilter(gxy); + gxy.x += 8u; + CurrFilter(gxy); + gxy.y += 8u; + CurrFilter(gxy); + gxy.x -= 8u; + CurrFilter(gxy); +} \ No newline at end of file diff --git a/Ryujinx.Graphics.OpenGL/Effects/Shaders/fsr_sharpening.glsl b/Ryujinx.Graphics.OpenGL/Effects/Shaders/fsr_sharpening.glsl new file mode 100644 index 00000000..d3b98729 --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/Shaders/fsr_sharpening.glsl @@ -0,0 +1,37 @@ +#version 430 core +precision mediump float; +layout (local_size_x = 64) in; +layout(rgba8, binding = 0, location=0) uniform image2D imgOutput; +layout( location=1 ) uniform sampler2D source; +layout( location=2 ) uniform float sharpening; + +#define A_GPU 1 +#define A_GLSL 1 +#include "ffx_a.h" + +#define FSR_RCAS_F 1 +AU4 con0; + +AF4 FsrRcasLoadF(ASU2 p) { return AF4(texelFetch(source, p, 0)); } +void FsrRcasInputF(inout AF1 r, inout AF1 g, inout AF1 b) {} + +#include "ffx_fsr1.h" + +void CurrFilter(AU2 pos) +{ + AF3 c; + FsrRcasF(c.r, c.g, c.b, pos, con0); + imageStore(imgOutput, ASU2(pos), AF4(c, 1)); +} + +void main() { + FsrRcasCon(con0, sharpening); + AU2 gxy = ARmp8x8(gl_LocalInvocationID.x) + AU2(gl_WorkGroupID.x << 4u, gl_WorkGroupID.y << 4u); + CurrFilter(gxy); + gxy.x += 8u; + CurrFilter(gxy); + gxy.y += 8u; + CurrFilter(gxy); + gxy.x -= 8u; + CurrFilter(gxy); +} \ No newline at end of file diff --git a/Ryujinx.Graphics.OpenGL/Effects/Shaders/fxaa.glsl b/Ryujinx.Graphics.OpenGL/Effects/Shaders/fxaa.glsl new file mode 100644 index 00000000..8bdcbca6 --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/Shaders/fxaa.glsl @@ -0,0 +1,1174 @@ +/*============================================================================ + + + NVIDIA FXAA 3.11 by TIMOTHY LOTTES + + +------------------------------------------------------------------------------ +COPYRIGHT (C) 2010, 2011 NVIDIA CORPORATION. ALL RIGHTS RESERVED. +------------------------------------------------------------------------------ +TO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE LAW, THIS SOFTWARE IS PROVIDED +*AS IS* AND NVIDIA AND ITS SUPPLIERS DISCLAIM ALL WARRANTIES, EITHER EXPRESS +OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, IMPLIED WARRANTIES OF +MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. IN NO EVENT SHALL NVIDIA +OR ITS SUPPLIERS BE LIABLE FOR ANY SPECIAL, INCIDENTAL, INDIRECT, OR +CONSEQUENTIAL DAMAGES WHATSOEVER (INCLUDING, WITHOUT LIMITATION, DAMAGES FOR +LOSS OF BUSINESS PROFITS, BUSINESS INTERRUPTION, LOSS OF BUSINESS INFORMATION, +OR ANY OTHER PECUNIARY LOSS) ARISING OUT OF THE USE OF OR INABILITY TO USE +THIS SOFTWARE, EVEN IF NVIDIA HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH +DAMAGES. + +------------------------------------------------------------------------------ + INTEGRATION CHECKLIST +------------------------------------------------------------------------------ +(1.) +In the shader source, setup defines for the desired configuration. +When providing multiple shaders (for different presets), +simply setup the defines differently in multiple files. +Example, + + #define FXAA_PC 1 + #define FXAA_HLSL_5 1 + #define FXAA_QUALITY__PRESET 12 + +Or, + + #define FXAA_360 1 + +Or, + + #define FXAA_PS3 1 + +Etc. + +(2.) +Then include this file, + + #include "Fxaa3_11.h" + +(3.) +Then call the FXAA pixel shader from within your desired shader. +Look at the FXAA Quality FxaaPixelShader() for docs on inputs. +As for FXAA 3.11 all inputs for all shaders are the same +to enable easy porting between platforms. + + return FxaaPixelShader(...); + +(4.) +Insure pass prior to FXAA outputs RGBL (see next section). +Or use, + + #define FXAA_GREEN_AS_LUMA 1 + +(5.) +Setup engine to provide the following constants +which are used in the FxaaPixelShader() inputs, + + FxaaFloat2 fxaaQualityRcpFrame, + FxaaFloat4 fxaaConsoleRcpFrameOpt, + FxaaFloat4 fxaaConsoleRcpFrameOpt2, + FxaaFloat4 fxaaConsole360RcpFrameOpt2, + FxaaFloat fxaaQualitySubpix, + FxaaFloat fxaaQualityEdgeThreshold, + FxaaFloat fxaaQualityEdgeThresholdMin, + FxaaFloat fxaaConsoleEdgeSharpness, + FxaaFloat fxaaConsoleEdgeThreshold, + FxaaFloat fxaaConsoleEdgeThresholdMin, + FxaaFloat4 fxaaConsole360ConstDir + +Look at the FXAA Quality FxaaPixelShader() for docs on inputs. + +(6.) +Have FXAA vertex shader run as a full screen triangle, +and output "pos" and "fxaaConsolePosPos" +such that inputs in the pixel shader provide, + + // {xy} = center of pixel + FxaaFloat2 pos, + + // {xy__} = upper left of pixel + // {__zw} = lower right of pixel + FxaaFloat4 fxaaConsolePosPos, + +(7.) +Insure the texture sampler(s) used by FXAA are set to bilinear filtering. + + +------------------------------------------------------------------------------ + INTEGRATION - RGBL AND COLORSPACE +------------------------------------------------------------------------------ +FXAA3 requires RGBL as input unless the following is set, + + #define FXAA_GREEN_AS_LUMA 1 + +In which case the engine uses green in place of luma, +and requires RGB input is in a non-linear colorspace. + +RGB should be LDR (low dynamic range). +Specifically do FXAA after tonemapping. + +RGB data as returned by a texture fetch can be non-linear, +or linear when FXAA_GREEN_AS_LUMA is not set. +Note an "sRGB format" texture counts as linear, +because the result of a texture fetch is linear data. +Regular "RGBA8" textures in the sRGB colorspace are non-linear. + +If FXAA_GREEN_AS_LUMA is not set, +luma must be stored in the alpha channel prior to running FXAA. +This luma should be in a perceptual space (could be gamma 2.0). +Example pass before FXAA where output is gamma 2.0 encoded, + + color.rgb = ToneMap(color.rgb); // linear color output + color.rgb = sqrt(color.rgb); // gamma 2.0 color output + return color; + +To use FXAA, + + color.rgb = ToneMap(color.rgb); // linear color output + color.rgb = sqrt(color.rgb); // gamma 2.0 color output + color.a = dot(color.rgb, FxaaFloat3(0.299, 0.587, 0.114)); // compute luma + return color; + +Another example where output is linear encoded, +say for instance writing to an sRGB formated render target, +where the render target does the conversion back to sRGB after blending, + + color.rgb = ToneMap(color.rgb); // linear color output + return color; + +To use FXAA, + + color.rgb = ToneMap(color.rgb); // linear color output + color.a = sqrt(dot(color.rgb, FxaaFloat3(0.299, 0.587, 0.114))); // compute luma + return color; + +Getting luma correct is required for the algorithm to work correctly. + + +------------------------------------------------------------------------------ + BEING LINEARLY CORRECT? +------------------------------------------------------------------------------ +Applying FXAA to a framebuffer with linear RGB color will look worse. +This is very counter intuitive, but happends to be true in this case. +The reason is because dithering artifacts will be more visiable +in a linear colorspace. + + +------------------------------------------------------------------------------ + COMPLEX INTEGRATION +------------------------------------------------------------------------------ +Q. What if the engine is blending into RGB before wanting to run FXAA? + +A. In the last opaque pass prior to FXAA, + have the pass write out luma into alpha. + Then blend into RGB only. + FXAA should be able to run ok + assuming the blending pass did not any add aliasing. + This should be the common case for particles and common blending passes. + +A. Or use FXAA_GREEN_AS_LUMA. + +============================================================================*/ + +#version 430 core + +layout(local_size_x = 16, local_size_y = 16) in; +layout(rgba8, binding = 0) uniform image2D imgOutput; + +uniform sampler2D inputTexture; +layout(location=0) uniform vec2 invResolution; + +#define FXAA_QUALITY__PRESET 12 +#define FXAA_GREEN_AS_LUMA 1 +#define FXAA_PC 1 +#define FXAA_GLSL_130 1 + + +/*============================================================================ + + INTEGRATION KNOBS + +/*==========================================================================*/ +#ifndef FXAA_PC + // + // FXAA Quality + // The high quality PC algorithm. + // + #define FXAA_PC 0 +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_GLSL_120 + #define FXAA_GLSL_120 0 +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_GLSL_130 + #define FXAA_GLSL_130 0 +#endif +/*==========================================================================*/ +#ifndef FXAA_GREEN_AS_LUMA + // + // For those using non-linear color, + // and either not able to get luma in alpha, or not wanting to, + // this enables FXAA to run using green as a proxy for luma. + // So with this enabled, no need to pack luma in alpha. + // + // This will turn off AA on anything which lacks some amount of green. + // Pure red and blue or combination of only R and B, will get no AA. + // + // Might want to lower the settings for both, + // fxaaConsoleEdgeThresholdMin + // fxaaQualityEdgeThresholdMin + // In order to insure AA does not get turned off on colors + // which contain a minor amount of green. + // + // 1 = On. + // 0 = Off. + // + #define FXAA_GREEN_AS_LUMA 0 +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_EARLY_EXIT + // + // Controls algorithm's early exit path. + // On PS3 turning this ON adds 2 cycles to the shader. + // On 360 turning this OFF adds 10ths of a millisecond to the shader. + // Turning this off on console will result in a more blurry image. + // So this defaults to on. + // + // 1 = On. + // 0 = Off. + // + #define FXAA_EARLY_EXIT 1 +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_DISCARD + // + // Only valid for PC OpenGL currently. + // Probably will not work when FXAA_GREEN_AS_LUMA = 1. + // + // 1 = Use discard on pixels which don't need AA. + // For APIs which enable concurrent TEX+ROP from same surface. + // 0 = Return unchanged color on pixels which don't need AA. + // + #define FXAA_DISCARD 0 +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_FAST_PIXEL_OFFSET + // + // Used for GLSL 120 only. + // + // 1 = GL API supports fast pixel offsets + // 0 = do not use fast pixel offsets + // + #ifdef GL_EXT_gpu_shader4 + #define FXAA_FAST_PIXEL_OFFSET 1 + #endif + #ifdef GL_NV_gpu_shader5 + #define FXAA_FAST_PIXEL_OFFSET 1 + #endif + #ifdef GL_ARB_gpu_shader5 + #define FXAA_FAST_PIXEL_OFFSET 1 + #endif + #ifndef FXAA_FAST_PIXEL_OFFSET + #define FXAA_FAST_PIXEL_OFFSET 0 + #endif +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_GATHER4_ALPHA + // + // 1 = API supports gather4 on alpha channel. + // 0 = API does not support gather4 on alpha channel. + // + #if (FXAA_HLSL_5 == 1) + #define FXAA_GATHER4_ALPHA 1 + #endif + #ifdef GL_ARB_gpu_shader5 + #define FXAA_GATHER4_ALPHA 1 + #endif + #ifdef GL_NV_gpu_shader5 + #define FXAA_GATHER4_ALPHA 1 + #endif + #ifndef FXAA_GATHER4_ALPHA + #define FXAA_GATHER4_ALPHA 0 + #endif +#endif + +/*============================================================================ + FXAA QUALITY - TUNING KNOBS +------------------------------------------------------------------------------ +NOTE the other tuning knobs are now in the shader function inputs! +============================================================================*/ +#ifndef FXAA_QUALITY__PRESET + // + // Choose the quality preset. + // This needs to be compiled into the shader as it effects code. + // Best option to include multiple presets is to + // in each shader define the preset, then include this file. + // + // OPTIONS + // ----------------------------------------------------------------------- + // 10 to 15 - default medium dither (10=fastest, 15=highest quality) + // 20 to 29 - less dither, more expensive (20=fastest, 29=highest quality) + // 39 - no dither, very expensive + // + // NOTES + // ----------------------------------------------------------------------- + // 12 = slightly faster then FXAA 3.9 and higher edge quality (default) + // 13 = about same speed as FXAA 3.9 and better than 12 + // 23 = closest to FXAA 3.9 visually and performance wise + // _ = the lowest digit is directly related to performance + // _ = the highest digit is directly related to style + // + #define FXAA_QUALITY__PRESET 12 +#endif + + +/*============================================================================ + + FXAA QUALITY - PRESETS + +============================================================================*/ + +/*============================================================================ + FXAA QUALITY - MEDIUM DITHER PRESETS +============================================================================*/ +#if (FXAA_QUALITY__PRESET == 10) + #define FXAA_QUALITY__PS 3 + #define FXAA_QUALITY__P0 1.5 + #define FXAA_QUALITY__P1 3.0 + #define FXAA_QUALITY__P2 12.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 11) + #define FXAA_QUALITY__PS 4 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 3.0 + #define FXAA_QUALITY__P3 12.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 12) + #define FXAA_QUALITY__PS 5 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 4.0 + #define FXAA_QUALITY__P4 12.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 13) + #define FXAA_QUALITY__PS 6 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 4.0 + #define FXAA_QUALITY__P5 12.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 14) + #define FXAA_QUALITY__PS 7 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 2.0 + #define FXAA_QUALITY__P5 4.0 + #define FXAA_QUALITY__P6 12.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 15) + #define FXAA_QUALITY__PS 8 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 2.0 + #define FXAA_QUALITY__P5 2.0 + #define FXAA_QUALITY__P6 4.0 + #define FXAA_QUALITY__P7 12.0 +#endif + +/*============================================================================ + FXAA QUALITY - LOW DITHER PRESETS +============================================================================*/ +#if (FXAA_QUALITY__PRESET == 20) + #define FXAA_QUALITY__PS 3 + #define FXAA_QUALITY__P0 1.5 + #define FXAA_QUALITY__P1 2.0 + #define FXAA_QUALITY__P2 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 21) + #define FXAA_QUALITY__PS 4 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 22) + #define FXAA_QUALITY__PS 5 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 23) + #define FXAA_QUALITY__PS 6 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 2.0 + #define FXAA_QUALITY__P5 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 24) + #define FXAA_QUALITY__PS 7 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 2.0 + #define FXAA_QUALITY__P5 3.0 + #define FXAA_QUALITY__P6 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 25) + #define FXAA_QUALITY__PS 8 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 2.0 + #define FXAA_QUALITY__P5 2.0 + #define FXAA_QUALITY__P6 4.0 + #define FXAA_QUALITY__P7 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 26) + #define FXAA_QUALITY__PS 9 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 2.0 + #define FXAA_QUALITY__P5 2.0 + #define FXAA_QUALITY__P6 2.0 + #define FXAA_QUALITY__P7 4.0 + #define FXAA_QUALITY__P8 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 27) + #define FXAA_QUALITY__PS 10 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 2.0 + #define FXAA_QUALITY__P5 2.0 + #define FXAA_QUALITY__P6 2.0 + #define FXAA_QUALITY__P7 2.0 + #define FXAA_QUALITY__P8 4.0 + #define FXAA_QUALITY__P9 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 28) + #define FXAA_QUALITY__PS 11 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 2.0 + #define FXAA_QUALITY__P5 2.0 + #define FXAA_QUALITY__P6 2.0 + #define FXAA_QUALITY__P7 2.0 + #define FXAA_QUALITY__P8 2.0 + #define FXAA_QUALITY__P9 4.0 + #define FXAA_QUALITY__P10 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY__PRESET == 29) + #define FXAA_QUALITY__PS 12 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.5 + #define FXAA_QUALITY__P2 2.0 + #define FXAA_QUALITY__P3 2.0 + #define FXAA_QUALITY__P4 2.0 + #define FXAA_QUALITY__P5 2.0 + #define FXAA_QUALITY__P6 2.0 + #define FXAA_QUALITY__P7 2.0 + #define FXAA_QUALITY__P8 2.0 + #define FXAA_QUALITY__P9 2.0 + #define FXAA_QUALITY__P10 4.0 + #define FXAA_QUALITY__P11 8.0 +#endif + +/*============================================================================ + FXAA QUALITY - EXTREME QUALITY +============================================================================*/ +#if (FXAA_QUALITY__PRESET == 39) + #define FXAA_QUALITY__PS 12 + #define FXAA_QUALITY__P0 1.0 + #define FXAA_QUALITY__P1 1.0 + #define FXAA_QUALITY__P2 1.0 + #define FXAA_QUALITY__P3 1.0 + #define FXAA_QUALITY__P4 1.0 + #define FXAA_QUALITY__P5 1.5 + #define FXAA_QUALITY__P6 2.0 + #define FXAA_QUALITY__P7 2.0 + #define FXAA_QUALITY__P8 2.0 + #define FXAA_QUALITY__P9 2.0 + #define FXAA_QUALITY__P10 4.0 + #define FXAA_QUALITY__P11 8.0 +#endif + + + +/*============================================================================ + + API PORTING + +============================================================================*/ +#if (FXAA_GLSL_120 == 1) || (FXAA_GLSL_130 == 1) + #define FxaaBool bool + #define FxaaDiscard discard + #define FxaaFloat float + #define FxaaFloat2 vec2 + #define FxaaFloat3 vec3 + #define FxaaFloat4 vec4 + #define FxaaHalf float + #define FxaaHalf2 vec2 + #define FxaaHalf3 vec3 + #define FxaaHalf4 vec4 + #define FxaaInt2 ivec2 + #define FxaaSat(x) clamp(x, 0.0, 1.0) + #define FxaaTex sampler2D +#else + #define FxaaBool bool + #define FxaaDiscard clip(-1) + #define FxaaFloat float + #define FxaaFloat2 float2 + #define FxaaFloat3 float3 + #define FxaaFloat4 float4 + #define FxaaHalf half + #define FxaaHalf2 half2 + #define FxaaHalf3 half3 + #define FxaaHalf4 half4 + #define FxaaSat(x) saturate(x) +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_GLSL_120 == 1) + // Requires, + // #version 120 + // And at least, + // #extension GL_EXT_gpu_shader4 : enable + // (or set FXAA_FAST_PIXEL_OFFSET 1 to work like DX9) + #define FxaaTexTop(t, p) texture2DLod(t, p, 0.0) + #if (FXAA_FAST_PIXEL_OFFSET == 1) + #define FxaaTexOff(t, p, o, r) texture2DLodOffset(t, p, 0.0, o) + #else + #define FxaaTexOff(t, p, o, r) texture2DLod(t, p + (o * r), 0.0) + #endif + #if (FXAA_GATHER4_ALPHA == 1) + // use #extension GL_ARB_gpu_shader5 : enable + #define FxaaTexAlpha4(t, p) textureGather(t, p, 3) + #define FxaaTexOffAlpha4(t, p, o) textureGatherOffset(t, p, o, 3) + #define FxaaTexGreen4(t, p) textureGather(t, p, 1) + #define FxaaTexOffGreen4(t, p, o) textureGatherOffset(t, p, o, 1) + #endif +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_GLSL_130 == 1) + // Requires "#version 130" or better + #define FxaaTexTop(t, p) textureLod(t, p, 0.0) + #define FxaaTexOff(t, p, o, r) textureLodOffset(t, p, 0.0, o) + #if (FXAA_GATHER4_ALPHA == 1) + // use #extension GL_ARB_gpu_shader5 : enable + #define FxaaTexAlpha4(t, p) textureGather(t, p, 3) + #define FxaaTexOffAlpha4(t, p, o) textureGatherOffset(t, p, o, 3) + #define FxaaTexGreen4(t, p) textureGather(t, p, 1) + #define FxaaTexOffGreen4(t, p, o) textureGatherOffset(t, p, o, 1) + #endif +#endif + + +/*============================================================================ + GREEN AS LUMA OPTION SUPPORT FUNCTION +============================================================================*/ +#if (FXAA_GREEN_AS_LUMA == 0) + FxaaFloat FxaaLuma(FxaaFloat4 rgba) { return rgba.w; } +#else + FxaaFloat FxaaLuma(FxaaFloat4 rgba) { return rgba.y; } +#endif + + + + +/*============================================================================ + + FXAA3 QUALITY - PC + +============================================================================*/ +#if (FXAA_PC == 1) +/*--------------------------------------------------------------------------*/ +FxaaFloat4 FxaaPixelShader( + // + // Use noperspective interpolation here (turn off perspective interpolation). + // {xy} = center of pixel + FxaaFloat2 pos, + // + // Used only for FXAA Console, and not used on the 360 version. + // Use noperspective interpolation here (turn off perspective interpolation). + // {xy__} = upper left of pixel + // {__zw} = lower right of pixel + FxaaFloat4 fxaaConsolePosPos, + // + // Input color texture. + // {rgb_} = color in linear or perceptual color space + // if (FXAA_GREEN_AS_LUMA == 0) + // {___a} = luma in perceptual color space (not linear) + FxaaTex tex, + // + // Only used on the optimized 360 version of FXAA Console. + // For everything but 360, just use the same input here as for "tex". + // For 360, same texture, just alias with a 2nd sampler. + // This sampler needs to have an exponent bias of -1. + FxaaTex fxaaConsole360TexExpBiasNegOne, + // + // Only used on the optimized 360 version of FXAA Console. + // For everything but 360, just use the same input here as for "tex". + // For 360, same texture, just alias with a 3nd sampler. + // This sampler needs to have an exponent bias of -2. + FxaaTex fxaaConsole360TexExpBiasNegTwo, + // + // Only used on FXAA Quality. + // This must be from a constant/uniform. + // {x_} = 1.0/screenWidthInPixels + // {_y} = 1.0/screenHeightInPixels + FxaaFloat2 fxaaQualityRcpFrame, + // + // Only used on FXAA Console. + // This must be from a constant/uniform. + // This effects sub-pixel AA quality and inversely sharpness. + // Where N ranges between, + // N = 0.50 (default) + // N = 0.33 (sharper) + // {x___} = -N/screenWidthInPixels + // {_y__} = -N/screenHeightInPixels + // {__z_} = N/screenWidthInPixels + // {___w} = N/screenHeightInPixels + FxaaFloat4 fxaaConsoleRcpFrameOpt, + // + // Only used on FXAA Console. + // Not used on 360, but used on PS3 and PC. + // This must be from a constant/uniform. + // {x___} = -2.0/screenWidthInPixels + // {_y__} = -2.0/screenHeightInPixels + // {__z_} = 2.0/screenWidthInPixels + // {___w} = 2.0/screenHeightInPixels + FxaaFloat4 fxaaConsoleRcpFrameOpt2, + // + // Only used on FXAA Console. + // Only used on 360 in place of fxaaConsoleRcpFrameOpt2. + // This must be from a constant/uniform. + // {x___} = 8.0/screenWidthInPixels + // {_y__} = 8.0/screenHeightInPixels + // {__z_} = -4.0/screenWidthInPixels + // {___w} = -4.0/screenHeightInPixels + FxaaFloat4 fxaaConsole360RcpFrameOpt2, + // + // Only used on FXAA Quality. + // This used to be the FXAA_QUALITY__SUBPIX define. + // It is here now to allow easier tuning. + // Choose the amount of sub-pixel aliasing removal. + // This can effect sharpness. + // 1.00 - upper limit (softer) + // 0.75 - default amount of filtering + // 0.50 - lower limit (sharper, less sub-pixel aliasing removal) + // 0.25 - almost off + // 0.00 - completely off + FxaaFloat fxaaQualitySubpix, + // + // Only used on FXAA Quality. + // This used to be the FXAA_QUALITY__EDGE_THRESHOLD define. + // It is here now to allow easier tuning. + // The minimum amount of local contrast required to apply algorithm. + // 0.333 - too little (faster) + // 0.250 - low quality + // 0.166 - default + // 0.125 - high quality + // 0.063 - overkill (slower) + FxaaFloat fxaaQualityEdgeThreshold, + // + // Only used on FXAA Quality. + // This used to be the FXAA_QUALITY__EDGE_THRESHOLD_MIN define. + // It is here now to allow easier tuning. + // Trims the algorithm from processing darks. + // 0.0833 - upper limit (default, the start of visible unfiltered edges) + // 0.0625 - high quality (faster) + // 0.0312 - visible limit (slower) + // Special notes when using FXAA_GREEN_AS_LUMA, + // Likely want to set this to zero. + // As colors that are mostly not-green + // will appear very dark in the green channel! + // Tune by looking at mostly non-green content, + // then start at zero and increase until aliasing is a problem. + FxaaFloat fxaaQualityEdgeThresholdMin, + // + // Only used on FXAA Console. + // This used to be the FXAA_CONSOLE__EDGE_SHARPNESS define. + // It is here now to allow easier tuning. + // This does not effect PS3, as this needs to be compiled in. + // Use FXAA_CONSOLE__PS3_EDGE_SHARPNESS for PS3. + // Due to the PS3 being ALU bound, + // there are only three safe values here: 2 and 4 and 8. + // These options use the shaders ability to a free *|/ by 2|4|8. + // For all other platforms can be a non-power of two. + // 8.0 is sharper (default!!!) + // 4.0 is softer + // 2.0 is really soft (good only for vector graphics inputs) + FxaaFloat fxaaConsoleEdgeSharpness, + // + // Only used on FXAA Console. + // This used to be the FXAA_CONSOLE__EDGE_THRESHOLD define. + // It is here now to allow easier tuning. + // This does not effect PS3, as this needs to be compiled in. + // Use FXAA_CONSOLE__PS3_EDGE_THRESHOLD for PS3. + // Due to the PS3 being ALU bound, + // there are only two safe values here: 1/4 and 1/8. + // These options use the shaders ability to a free *|/ by 2|4|8. + // The console setting has a different mapping than the quality setting. + // Other platforms can use other values. + // 0.125 leaves less aliasing, but is softer (default!!!) + // 0.25 leaves more aliasing, and is sharper + FxaaFloat fxaaConsoleEdgeThreshold, + // + // Only used on FXAA Console. + // This used to be the FXAA_CONSOLE__EDGE_THRESHOLD_MIN define. + // It is here now to allow easier tuning. + // Trims the algorithm from processing darks. + // The console setting has a different mapping than the quality setting. + // This only applies when FXAA_EARLY_EXIT is 1. + // This does not apply to PS3, + // PS3 was simplified to avoid more shader instructions. + // 0.06 - faster but more aliasing in darks + // 0.05 - default + // 0.04 - slower and less aliasing in darks + // Special notes when using FXAA_GREEN_AS_LUMA, + // Likely want to set this to zero. + // As colors that are mostly not-green + // will appear very dark in the green channel! + // Tune by looking at mostly non-green content, + // then start at zero and increase until aliasing is a problem. + FxaaFloat fxaaConsoleEdgeThresholdMin, + // + // Extra constants for 360 FXAA Console only. + // Use zeros or anything else for other platforms. + // These must be in physical constant registers and NOT immedates. + // Immedates will result in compiler un-optimizing. + // {xyzw} = float4(1.0, -1.0, 0.25, -0.25) + FxaaFloat4 fxaaConsole360ConstDir +) { +/*--------------------------------------------------------------------------*/ + FxaaFloat2 posM; + posM.x = pos.x; + posM.y = pos.y; + #if (FXAA_GATHER4_ALPHA == 1) + #if (FXAA_DISCARD == 0) + FxaaFloat4 rgbyM = FxaaTexTop(tex, posM); + #if (FXAA_GREEN_AS_LUMA == 0) + #define lumaM rgbyM.w + #else + #define lumaM rgbyM.y + #endif + #endif + #if (FXAA_GREEN_AS_LUMA == 0) + FxaaFloat4 luma4A = FxaaTexAlpha4(tex, posM); + FxaaFloat4 luma4B = FxaaTexOffAlpha4(tex, posM, FxaaInt2(-1, -1)); + #else + FxaaFloat4 luma4A = FxaaTexGreen4(tex, posM); + FxaaFloat4 luma4B = FxaaTexOffGreen4(tex, posM, FxaaInt2(-1, -1)); + #endif + #if (FXAA_DISCARD == 1) + #define lumaM luma4A.w + #endif + #define lumaE luma4A.z + #define lumaS luma4A.x + #define lumaSE luma4A.y + #define lumaNW luma4B.w + #define lumaN luma4B.z + #define lumaW luma4B.x + #else + FxaaFloat4 rgbyM = FxaaTexTop(tex, posM); + #if (FXAA_GREEN_AS_LUMA == 0) + #define lumaM rgbyM.w + #else + #define lumaM rgbyM.y + #endif + FxaaFloat lumaS = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2( 0, 1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaE = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2( 1, 0), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaN = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2( 0,-1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaW = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2(-1, 0), fxaaQualityRcpFrame.xy)); + #endif +/*--------------------------------------------------------------------------*/ + FxaaFloat maxSM = max(lumaS, lumaM); + FxaaFloat minSM = min(lumaS, lumaM); + FxaaFloat maxESM = max(lumaE, maxSM); + FxaaFloat minESM = min(lumaE, minSM); + FxaaFloat maxWN = max(lumaN, lumaW); + FxaaFloat minWN = min(lumaN, lumaW); + FxaaFloat rangeMax = max(maxWN, maxESM); + FxaaFloat rangeMin = min(minWN, minESM); + FxaaFloat rangeMaxScaled = rangeMax * fxaaQualityEdgeThreshold; + FxaaFloat range = rangeMax - rangeMin; + FxaaFloat rangeMaxClamped = max(fxaaQualityEdgeThresholdMin, rangeMaxScaled); + FxaaBool earlyExit = range < rangeMaxClamped; +/*--------------------------------------------------------------------------*/ + if(earlyExit) + #if (FXAA_DISCARD == 1) + FxaaDiscard; + #else + return rgbyM; + #endif +/*--------------------------------------------------------------------------*/ + #if (FXAA_GATHER4_ALPHA == 0) + FxaaFloat lumaNW = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2(-1,-1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaSE = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2( 1, 1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaNE = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2( 1,-1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaSW = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2(-1, 1), fxaaQualityRcpFrame.xy)); + #else + FxaaFloat lumaNE = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2(1, -1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaSW = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2(-1, 1), fxaaQualityRcpFrame.xy)); + #endif +/*--------------------------------------------------------------------------*/ + FxaaFloat lumaNS = lumaN + lumaS; + FxaaFloat lumaWE = lumaW + lumaE; + FxaaFloat subpixRcpRange = 1.0/range; + FxaaFloat subpixNSWE = lumaNS + lumaWE; + FxaaFloat edgeHorz1 = (-2.0 * lumaM) + lumaNS; + FxaaFloat edgeVert1 = (-2.0 * lumaM) + lumaWE; +/*--------------------------------------------------------------------------*/ + FxaaFloat lumaNESE = lumaNE + lumaSE; + FxaaFloat lumaNWNE = lumaNW + lumaNE; + FxaaFloat edgeHorz2 = (-2.0 * lumaE) + lumaNESE; + FxaaFloat edgeVert2 = (-2.0 * lumaN) + lumaNWNE; +/*--------------------------------------------------------------------------*/ + FxaaFloat lumaNWSW = lumaNW + lumaSW; + FxaaFloat lumaSWSE = lumaSW + lumaSE; + FxaaFloat edgeHorz4 = (abs(edgeHorz1) * 2.0) + abs(edgeHorz2); + FxaaFloat edgeVert4 = (abs(edgeVert1) * 2.0) + abs(edgeVert2); + FxaaFloat edgeHorz3 = (-2.0 * lumaW) + lumaNWSW; + FxaaFloat edgeVert3 = (-2.0 * lumaS) + lumaSWSE; + FxaaFloat edgeHorz = abs(edgeHorz3) + edgeHorz4; + FxaaFloat edgeVert = abs(edgeVert3) + edgeVert4; +/*--------------------------------------------------------------------------*/ + FxaaFloat subpixNWSWNESE = lumaNWSW + lumaNESE; + FxaaFloat lengthSign = fxaaQualityRcpFrame.x; + FxaaBool horzSpan = edgeHorz >= edgeVert; + FxaaFloat subpixA = subpixNSWE * 2.0 + subpixNWSWNESE; +/*--------------------------------------------------------------------------*/ + if(!horzSpan) lumaN = lumaW; + if(!horzSpan) lumaS = lumaE; + if(horzSpan) lengthSign = fxaaQualityRcpFrame.y; + FxaaFloat subpixB = (subpixA * (1.0/12.0)) - lumaM; +/*--------------------------------------------------------------------------*/ + FxaaFloat gradientN = lumaN - lumaM; + FxaaFloat gradientS = lumaS - lumaM; + FxaaFloat lumaNN = lumaN + lumaM; + FxaaFloat lumaSS = lumaS + lumaM; + FxaaBool pairN = abs(gradientN) >= abs(gradientS); + FxaaFloat gradient = max(abs(gradientN), abs(gradientS)); + if(pairN) lengthSign = -lengthSign; + FxaaFloat subpixC = FxaaSat(abs(subpixB) * subpixRcpRange); +/*--------------------------------------------------------------------------*/ + FxaaFloat2 posB; + posB.x = posM.x; + posB.y = posM.y; + FxaaFloat2 offNP; + offNP.x = (!horzSpan) ? 0.0 : fxaaQualityRcpFrame.x; + offNP.y = ( horzSpan) ? 0.0 : fxaaQualityRcpFrame.y; + if(!horzSpan) posB.x += lengthSign * 0.5; + if( horzSpan) posB.y += lengthSign * 0.5; +/*--------------------------------------------------------------------------*/ + FxaaFloat2 posN; + posN.x = posB.x - offNP.x * FXAA_QUALITY__P0; + posN.y = posB.y - offNP.y * FXAA_QUALITY__P0; + FxaaFloat2 posP; + posP.x = posB.x + offNP.x * FXAA_QUALITY__P0; + posP.y = posB.y + offNP.y * FXAA_QUALITY__P0; + FxaaFloat subpixD = ((-2.0)*subpixC) + 3.0; + FxaaFloat lumaEndN = FxaaLuma(FxaaTexTop(tex, posN)); + FxaaFloat subpixE = subpixC * subpixC; + FxaaFloat lumaEndP = FxaaLuma(FxaaTexTop(tex, posP)); +/*--------------------------------------------------------------------------*/ + if(!pairN) lumaNN = lumaSS; + FxaaFloat gradientScaled = gradient * 1.0/4.0; + FxaaFloat lumaMM = lumaM - lumaNN * 0.5; + FxaaFloat subpixF = subpixD * subpixE; + FxaaBool lumaMLTZero = lumaMM < 0.0; +/*--------------------------------------------------------------------------*/ + lumaEndN -= lumaNN * 0.5; + lumaEndP -= lumaNN * 0.5; + FxaaBool doneN = abs(lumaEndN) >= gradientScaled; + FxaaBool doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P1; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P1; + FxaaBool doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P1; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P1; +/*--------------------------------------------------------------------------*/ + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P2; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P2; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P2; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P2; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY__PS > 3) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P3; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P3; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P3; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P3; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY__PS > 4) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P4; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P4; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P4; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P4; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY__PS > 5) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P5; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P5; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P5; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P5; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY__PS > 6) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P6; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P6; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P6; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P6; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY__PS > 7) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P7; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P7; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P7; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P7; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY__PS > 8) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P8; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P8; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P8; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P8; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY__PS > 9) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P9; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P9; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P9; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P9; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY__PS > 10) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P10; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P10; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P10; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P10; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY__PS > 11) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P11; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P11; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P11; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P11; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY__PS > 12) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY__P12; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY__P12; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY__P12; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY__P12; +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } +/*--------------------------------------------------------------------------*/ + FxaaFloat dstN = posM.x - posN.x; + FxaaFloat dstP = posP.x - posM.x; + if(!horzSpan) dstN = posM.y - posN.y; + if(!horzSpan) dstP = posP.y - posM.y; +/*--------------------------------------------------------------------------*/ + FxaaBool goodSpanN = (lumaEndN < 0.0) != lumaMLTZero; + FxaaFloat spanLength = (dstP + dstN); + FxaaBool goodSpanP = (lumaEndP < 0.0) != lumaMLTZero; + FxaaFloat spanLengthRcp = 1.0/spanLength; +/*--------------------------------------------------------------------------*/ + FxaaBool directionN = dstN < dstP; + FxaaFloat dst = min(dstN, dstP); + FxaaBool goodSpan = directionN ? goodSpanN : goodSpanP; + FxaaFloat subpixG = subpixF * subpixF; + FxaaFloat pixelOffset = (dst * (-spanLengthRcp)) + 0.5; + FxaaFloat subpixH = subpixG * fxaaQualitySubpix; +/*--------------------------------------------------------------------------*/ + FxaaFloat pixelOffsetGood = goodSpan ? pixelOffset : 0.0; + FxaaFloat pixelOffsetSubpix = max(pixelOffsetGood, subpixH); + if(!horzSpan) posM.x += pixelOffsetSubpix * lengthSign; + if( horzSpan) posM.y += pixelOffsetSubpix * lengthSign; + #if (FXAA_DISCARD == 1) + return FxaaTexTop(tex, posM); + #else + return FxaaFloat4(FxaaTexTop(tex, posM).xyz, lumaM); + #endif +} +/*==========================================================================*/ +#endif + +vec4 mainImage(vec2 fragCoord) +{ + vec2 rcpFrame = 1./invResolution.xy; + vec2 uv2 = fragCoord.xy / invResolution.xy; + + float fxaaQualitySubpix = 0.75; // [0..1], default 0.75 + float fxaaQualityEdgeThreshold = 0.166; // [0.125..0.33], default 0.166 + float fxaaQualityEdgeThresholdMin = 0.02;//0.0625; // ? + vec4 dummy4 = vec4(0.0,0.0,0.0,0.0); + float dummy1 = 0.0; + + vec4 col = FxaaPixelShader(uv2, dummy4, + inputTexture, inputTexture, inputTexture, + rcpFrame, dummy4, dummy4, dummy4, + fxaaQualitySubpix, fxaaQualityEdgeThreshold, + fxaaQualityEdgeThresholdMin, + dummy1, dummy1, dummy1, dummy4); + + vec4 fragColor = vec4( col.xyz, 1. ); + + return fragColor; +} + +void main() +{ + ivec2 loc = ivec2(gl_GlobalInvocationID.x * 4, gl_GlobalInvocationID.y * 4); + for(int i = 0; i < 4; i++) + { + for(int j = 0; j < 4; j++) + { + ivec2 texelCoord = ivec2(loc.x + i, loc.y + j); + vec4 outColor = mainImage(texelCoord + vec2(0.5)); + imageStore(imgOutput, texelCoord, outColor); + } + } +} diff --git a/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa.hlsl b/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa.hlsl new file mode 100644 index 00000000..2201f78c --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa.hlsl @@ -0,0 +1,1361 @@ +/** + * Copyright (C) 2013 Jorge Jimenez (jorge@iryoku.com) + * Copyright (C) 2013 Jose I. Echevarria (joseignacioechevarria@gmail.com) + * Copyright (C) 2013 Belen Masia (bmasia@unizar.es) + * Copyright (C) 2013 Fernando Navarro (fernandn@microsoft.com) + * Copyright (C) 2013 Diego Gutierrez (diegog@unizar.es) + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * this software and associated documentation files (the "Software"), to deal in + * the Software without restriction, including without limitation the rights to + * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies + * of the Software, and to permit persons to whom the Software is furnished to + * do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. As clarification, there + * is no requirement that the copyright notice and permission be included in + * binary distributions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + + +/** + * _______ ___ ___ ___ ___ + * / || \/ | / \ / \ + * | (---- | \ / | / ^ \ / ^ \ + * \ \ | |\/| | / /_\ \ / /_\ \ + * ----) | | | | | / _____ \ / _____ \ + * |_______/ |__| |__| /__/ \__\ /__/ \__\ + * + * E N H A N C E D + * S U B P I X E L M O R P H O L O G I C A L A N T I A L I A S I N G + * + * http://www.iryoku.com/smaa/ + * + * Hi, welcome aboard! + * + * Here you'll find instructions to get the shader up and running as fast as + * possible. + * + * IMPORTANTE NOTICE: when updating, remember to update both this file and the + * precomputed textures! They may change from version to version. + * + * The shader has three passes, chained together as follows: + * + * |input|------------------� + * v | + * [ SMAA*EdgeDetection ] | + * v | + * |edgesTex| | + * v | + * [ SMAABlendingWeightCalculation ] | + * v | + * |blendTex| | + * v | + * [ SMAANeighborhoodBlending ] <------� + * v + * |output| + * + * Note that each [pass] has its own vertex and pixel shader. Remember to use + * oversized triangles instead of quads to avoid overshading along the + * diagonal. + * + * You've three edge detection methods to choose from: luma, color or depth. + * They represent different quality/performance and anti-aliasing/sharpness + * tradeoffs, so our recommendation is for you to choose the one that best + * suits your particular scenario: + * + * - Depth edge detection is usually the fastest but it may miss some edges. + * + * - Luma edge detection is usually more expensive than depth edge detection, + * but catches visible edges that depth edge detection can miss. + * + * - Color edge detection is usually the most expensive one but catches + * chroma-only edges. + * + * For quickstarters: just use luma edge detection. + * + * The general advice is to not rush the integration process and ensure each + * step is done correctly (don't try to integrate SMAA T2x with predicated edge + * detection from the start!). Ok then, let's go! + * + * 1. The first step is to create two RGBA temporal render targets for holding + * |edgesTex| and |blendTex|. + * + * In DX10 or DX11, you can use a RG render target for the edges texture. + * In the case of NVIDIA GPUs, using RG render targets seems to actually be + * slower. + * + * On the Xbox 360, you can use the same render target for resolving both + * |edgesTex| and |blendTex|, as they aren't needed simultaneously. + * + * 2. Both temporal render targets |edgesTex| and |blendTex| must be cleared + * each frame. Do not forget to clear the alpha channel! + * + * 3. The next step is loading the two supporting precalculated textures, + * 'areaTex' and 'searchTex'. You'll find them in the 'Textures' folder as + * C++ headers, and also as regular DDS files. They'll be needed for the + * 'SMAABlendingWeightCalculation' pass. + * + * If you use the C++ headers, be sure to load them in the format specified + * inside of them. + * + * You can also compress 'areaTex' and 'searchTex' using BC5 and BC4 + * respectively, if you have that option in your content processor pipeline. + * When compressing then, you get a non-perceptible quality decrease, and a + * marginal performance increase. + * + * 4. All samplers must be set to linear filtering and clamp. + * + * After you get the technique working, remember that 64-bit inputs have + * half-rate linear filtering on GCN. + * + * If SMAA is applied to 64-bit color buffers, switching to point filtering + * when accesing them will increase the performance. Search for + * 'SMAASamplePoint' to see which textures may benefit from point + * filtering, and where (which is basically the color input in the edge + * detection and resolve passes). + * + * 5. All texture reads and buffer writes must be non-sRGB, with the exception + * of the input read and the output write in + * 'SMAANeighborhoodBlending' (and only in this pass!). If sRGB reads in + * this last pass are not possible, the technique will work anyway, but + * will perform antialiasing in gamma space. + * + * IMPORTANT: for best results the input read for the color/luma edge + * detection should *NOT* be sRGB. + * + * 6. Before including SMAA.h you'll have to setup the render target metrics, + * the target and any optional configuration defines. Optionally you can + * use a preset. + * + * You have the following targets available: + * SMAA_HLSL_3 + * SMAA_HLSL_4 + * SMAA_HLSL_4_1 + * SMAA_GLSL_3 * + * SMAA_GLSL_4 * + * + * * (See SMAA_INCLUDE_VS and SMAA_INCLUDE_PS below). + * + * And four presets: + * SMAA_PRESET_LOW (%60 of the quality) + * SMAA_PRESET_MEDIUM (%80 of the quality) + * SMAA_PRESET_HIGH (%95 of the quality) + * SMAA_PRESET_ULTRA (%99 of the quality) + * + * For example: + * #define SMAA_RT_METRICS float4(1.0 / 1280.0, 1.0 / 720.0, 1280.0, 720.0) + * #define SMAA_HLSL_4 + * #define SMAA_PRESET_HIGH + * #include "SMAA.h" + * + * Note that SMAA_RT_METRICS doesn't need to be a macro, it can be a + * uniform variable. The code is designed to minimize the impact of not + * using a constant value, but it is still better to hardcode it. + * + * Depending on how you encoded 'areaTex' and 'searchTex', you may have to + * add (and customize) the following defines before including SMAA.h: + * #define SMAA_AREATEX_SELECT(sample) sample.rg + * #define SMAA_SEARCHTEX_SELECT(sample) sample.r + * + * If your engine is already using porting macros, you can define + * SMAA_CUSTOM_SL, and define the porting functions by yourself. + * + * 7. Then, you'll have to setup the passes as indicated in the scheme above. + * You can take a look into SMAA.fx, to see how we did it for our demo. + * Checkout the function wrappers, you may want to copy-paste them! + * + * 8. It's recommended to validate the produced |edgesTex| and |blendTex|. + * You can use a screenshot from your engine to compare the |edgesTex| + * and |blendTex| produced inside of the engine with the results obtained + * with the reference demo. + * + * 9. After you get the last pass to work, it's time to optimize. You'll have + * to initialize a stencil buffer in the first pass (discard is already in + * the code), then mask execution by using it the second pass. The last + * pass should be executed in all pixels. + * + * + * After this point you can choose to enable predicated thresholding, + * temporal supersampling and motion blur integration: + * + * a) If you want to use predicated thresholding, take a look into + * SMAA_PREDICATION; you'll need to pass an extra texture in the edge + * detection pass. + * + * b) If you want to enable temporal supersampling (SMAA T2x): + * + * 1. The first step is to render using subpixel jitters. I won't go into + * detail, but it's as simple as moving each vertex position in the + * vertex shader, you can check how we do it in our DX10 demo. + * + * 2. Then, you must setup the temporal resolve. You may want to take a look + * into SMAAResolve for resolving 2x modes. After you get it working, you'll + * probably see ghosting everywhere. But fear not, you can enable the + * CryENGINE temporal reprojection by setting the SMAA_REPROJECTION macro. + * Check out SMAA_DECODE_VELOCITY if your velocity buffer is encoded. + * + * 3. The next step is to apply SMAA to each subpixel jittered frame, just as + * done for 1x. + * + * 4. At this point you should already have something usable, but for best + * results the proper area textures must be set depending on current jitter. + * For this, the parameter 'subsampleIndices' of + * 'SMAABlendingWeightCalculationPS' must be set as follows, for our T2x + * mode: + * + * @SUBSAMPLE_INDICES + * + * | S# | Camera Jitter | subsampleIndices | + * +----+------------------+---------------------+ + * | 0 | ( 0.25, -0.25) | float4(1, 1, 1, 0) | + * | 1 | (-0.25, 0.25) | float4(2, 2, 2, 0) | + * + * These jitter positions assume a bottom-to-top y axis. S# stands for the + * sample number. + * + * More information about temporal supersampling here: + * http://iryoku.com/aacourse/downloads/13-Anti-Aliasing-Methods-in-CryENGINE-3.pdf + * + * c) If you want to enable spatial multisampling (SMAA S2x): + * + * 1. The scene must be rendered using MSAA 2x. The MSAA 2x buffer must be + * created with: + * - DX10: see below (*) + * - DX10.1: D3D10_STANDARD_MULTISAMPLE_PATTERN or + * - DX11: D3D11_STANDARD_MULTISAMPLE_PATTERN + * + * This allows to ensure that the subsample order matches the table in + * @SUBSAMPLE_INDICES. + * + * (*) In the case of DX10, we refer the reader to: + * - SMAA::detectMSAAOrder and + * - SMAA::msaaReorder + * + * These functions allow to match the standard multisample patterns by + * detecting the subsample order for a specific GPU, and reordering + * them appropriately. + * + * 2. A shader must be run to output each subsample into a separate buffer + * (DX10 is required). You can use SMAASeparate for this purpose, or just do + * it in an existing pass (for example, in the tone mapping pass, which has + * the advantage of feeding tone mapped subsamples to SMAA, which will yield + * better results). + * + * 3. The full SMAA 1x pipeline must be run for each separated buffer, storing + * the results in the final buffer. The second run should alpha blend with + * the existing final buffer using a blending factor of 0.5. + * 'subsampleIndices' must be adjusted as in the SMAA T2x case (see point + * b). + * + * d) If you want to enable temporal supersampling on top of SMAA S2x + * (which actually is SMAA 4x): + * + * 1. SMAA 4x consists on temporally jittering SMAA S2x, so the first step is + * to calculate SMAA S2x for current frame. In this case, 'subsampleIndices' + * must be set as follows: + * + * | F# | S# | Camera Jitter | Net Jitter | subsampleIndices | + * +----+----+--------------------+-------------------+----------------------+ + * | 0 | 0 | ( 0.125, 0.125) | ( 0.375, -0.125) | float4(5, 3, 1, 3) | + * | 0 | 1 | ( 0.125, 0.125) | (-0.125, 0.375) | float4(4, 6, 2, 3) | + * +----+----+--------------------+-------------------+----------------------+ + * | 1 | 2 | (-0.125, -0.125) | ( 0.125, -0.375) | float4(3, 5, 1, 4) | + * | 1 | 3 | (-0.125, -0.125) | (-0.375, 0.125) | float4(6, 4, 2, 4) | + * + * These jitter positions assume a bottom-to-top y axis. F# stands for the + * frame number. S# stands for the sample number. + * + * 2. After calculating SMAA S2x for current frame (with the new subsample + * indices), previous frame must be reprojected as in SMAA T2x mode (see + * point b). + * + * e) If motion blur is used, you may want to do the edge detection pass + * together with motion blur. This has two advantages: + * + * 1. Pixels under heavy motion can be omitted from the edge detection process. + * For these pixels we can just store "no edge", as motion blur will take + * care of them. + * 2. The center pixel tap is reused. + * + * Note that in this case depth testing should be used instead of stenciling, + * as we have to write all the pixels in the motion blur pass. + * + * That's it! + */ + +//----------------------------------------------------------------------------- +// SMAA Presets + +/** + * Note that if you use one of these presets, the following configuration + * macros will be ignored if set in the "Configurable Defines" section. + */ + +#if defined(SMAA_PRESET_LOW) +#define SMAA_THRESHOLD 0.15 +#define SMAA_MAX_SEARCH_STEPS 4 +#define SMAA_DISABLE_DIAG_DETECTION +#define SMAA_DISABLE_CORNER_DETECTION +#elif defined(SMAA_PRESET_MEDIUM) +#define SMAA_THRESHOLD 0.1 +#define SMAA_MAX_SEARCH_STEPS 8 +#define SMAA_DISABLE_DIAG_DETECTION +#define SMAA_DISABLE_CORNER_DETECTION +#elif defined(SMAA_PRESET_HIGH) +#define SMAA_THRESHOLD 0.1 +#define SMAA_MAX_SEARCH_STEPS 16 +#define SMAA_MAX_SEARCH_STEPS_DIAG 8 +#define SMAA_CORNER_ROUNDING 25 +#elif defined(SMAA_PRESET_ULTRA) +#define SMAA_THRESHOLD 0.05 +#define SMAA_MAX_SEARCH_STEPS 32 +#define SMAA_MAX_SEARCH_STEPS_DIAG 16 +#define SMAA_CORNER_ROUNDING 25 +#endif + +//----------------------------------------------------------------------------- +// Configurable Defines + +/** + * SMAA_THRESHOLD specifies the threshold or sensitivity to edges. + * Lowering this value you will be able to detect more edges at the expense of + * performance. + * + * Range: [0, 0.5] + * 0.1 is a reasonable value, and allows to catch most visible edges. + * 0.05 is a rather overkill value, that allows to catch 'em all. + * + * If temporal supersampling is used, 0.2 could be a reasonable value, as low + * contrast edges are properly filtered by just 2x. + */ +#ifndef SMAA_THRESHOLD +#define SMAA_THRESHOLD 0.1 +#endif + +/** + * SMAA_DEPTH_THRESHOLD specifies the threshold for depth edge detection. + * + * Range: depends on the depth range of the scene. + */ +#ifndef SMAA_DEPTH_THRESHOLD +#define SMAA_DEPTH_THRESHOLD (0.1 * SMAA_THRESHOLD) +#endif + +/** + * SMAA_MAX_SEARCH_STEPS specifies the maximum steps performed in the + * horizontal/vertical pattern searches, at each side of the pixel. + * + * In number of pixels, it's actually the double. So the maximum line length + * perfectly handled by, for example 16, is 64 (by perfectly, we meant that + * longer lines won't look as good, but still antialiased). + * + * Range: [0, 112] + */ +#ifndef SMAA_MAX_SEARCH_STEPS +#define SMAA_MAX_SEARCH_STEPS 16 +#endif + +/** + * SMAA_MAX_SEARCH_STEPS_DIAG specifies the maximum steps performed in the + * diagonal pattern searches, at each side of the pixel. In this case we jump + * one pixel at time, instead of two. + * + * Range: [0, 20] + * + * On high-end machines it is cheap (between a 0.8x and 0.9x slower for 16 + * steps), but it can have a significant impact on older machines. + * + * Define SMAA_DISABLE_DIAG_DETECTION to disable diagonal processing. + */ +#ifndef SMAA_MAX_SEARCH_STEPS_DIAG +#define SMAA_MAX_SEARCH_STEPS_DIAG 8 +#endif + +/** + * SMAA_CORNER_ROUNDING specifies how much sharp corners will be rounded. + * + * Range: [0, 100] + * + * Define SMAA_DISABLE_CORNER_DETECTION to disable corner processing. + */ +#ifndef SMAA_CORNER_ROUNDING +#define SMAA_CORNER_ROUNDING 25 +#endif + +/** + * If there is an neighbor edge that has SMAA_LOCAL_CONTRAST_FACTOR times + * bigger contrast than current edge, current edge will be discarded. + * + * This allows to eliminate spurious crossing edges, and is based on the fact + * that, if there is too much contrast in a direction, that will hide + * perceptually contrast in the other neighbors. + */ +#ifndef SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR +#define SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR 2.0 +#endif + +/** + * Predicated thresholding allows to better preserve texture details and to + * improve performance, by decreasing the number of detected edges using an + * additional buffer like the light accumulation buffer, object ids or even the + * depth buffer (the depth buffer usage may be limited to indoor or short range + * scenes). + * + * It locally decreases the luma or color threshold if an edge is found in an + * additional buffer (so the global threshold can be higher). + * + * This method was developed by Playstation EDGE MLAA team, and used in + * Killzone 3, by using the light accumulation buffer. More information here: + * http://iryoku.com/aacourse/downloads/06-MLAA-on-PS3.pptx + */ +#ifndef SMAA_PREDICATION +#define SMAA_PREDICATION 0 +#endif + +/** + * Threshold to be used in the additional predication buffer. + * + * Range: depends on the input, so you'll have to find the magic number that + * works for you. + */ +#ifndef SMAA_PREDICATION_THRESHOLD +#define SMAA_PREDICATION_THRESHOLD 0.01 +#endif + +/** + * How much to scale the global threshold used for luma or color edge + * detection when using predication. + * + * Range: [1, 5] + */ +#ifndef SMAA_PREDICATION_SCALE +#define SMAA_PREDICATION_SCALE 2.0 +#endif + +/** + * How much to locally decrease the threshold. + * + * Range: [0, 1] + */ +#ifndef SMAA_PREDICATION_STRENGTH +#define SMAA_PREDICATION_STRENGTH 0.4 +#endif + +/** + * Temporal reprojection allows to remove ghosting artifacts when using + * temporal supersampling. We use the CryEngine 3 method which also introduces + * velocity weighting. This feature is of extreme importance for totally + * removing ghosting. More information here: + * http://iryoku.com/aacourse/downloads/13-Anti-Aliasing-Methods-in-CryENGINE-3.pdf + * + * Note that you'll need to setup a velocity buffer for enabling reprojection. + * For static geometry, saving the previous depth buffer is a viable + * alternative. + */ +#ifndef SMAA_REPROJECTION +#define SMAA_REPROJECTION 0 +#endif + +/** + * SMAA_REPROJECTION_WEIGHT_SCALE controls the velocity weighting. It allows to + * remove ghosting trails behind the moving object, which are not removed by + * just using reprojection. Using low values will exhibit ghosting, while using + * high values will disable temporal supersampling under motion. + * + * Behind the scenes, velocity weighting removes temporal supersampling when + * the velocity of the subsamples differs (meaning they are different objects). + * + * Range: [0, 80] + */ +#ifndef SMAA_REPROJECTION_WEIGHT_SCALE +#define SMAA_REPROJECTION_WEIGHT_SCALE 30.0 +#endif + +/** + * On some compilers, discard cannot be used in vertex shaders. Thus, they need + * to be compiled separately. + */ +#ifndef SMAA_INCLUDE_VS +#define SMAA_INCLUDE_VS 1 +#endif +#ifndef SMAA_INCLUDE_PS +#define SMAA_INCLUDE_PS 1 +#endif + +//----------------------------------------------------------------------------- +// Texture Access Defines + +#ifndef SMAA_AREATEX_SELECT +#if defined(SMAA_HLSL_3) +#define SMAA_AREATEX_SELECT(sample) sample.ra +#else +#define SMAA_AREATEX_SELECT(sample) sample.rg +#endif +#endif + +#ifndef SMAA_SEARCHTEX_SELECT +#define SMAA_SEARCHTEX_SELECT(sample) sample.r +#endif + +#ifndef SMAA_DECODE_VELOCITY +#define SMAA_DECODE_VELOCITY(sample) sample.rg +#endif + +//----------------------------------------------------------------------------- +// Non-Configurable Defines + +#define SMAA_AREATEX_MAX_DISTANCE 16 +#define SMAA_AREATEX_MAX_DISTANCE_DIAG 20 +#define SMAA_AREATEX_PIXEL_SIZE (1.0 / float2(160.0, 560.0)) +#define SMAA_AREATEX_SUBTEX_SIZE (1.0 / 7.0) +#define SMAA_SEARCHTEX_SIZE float2(66.0, 33.0) +#define SMAA_SEARCHTEX_PACKED_SIZE float2(64.0, 16.0) +#define SMAA_CORNER_ROUNDING_NORM (float(SMAA_CORNER_ROUNDING) / 100.0) + +//----------------------------------------------------------------------------- +// Porting Functions + +#if defined(SMAA_HLSL_3) +#define SMAATexture2D(tex) sampler2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) tex2Dlod(tex, float4(coord, 0.0, 0.0)) +#define SMAASampleLevelZeroPoint(tex, coord) tex2Dlod(tex, float4(coord, 0.0, 0.0)) +#define SMAASampleLevelZeroOffset(tex, coord, offset) tex2Dlod(tex, float4(coord + offset * SMAA_RT_METRICS.xy, 0.0, 0.0)) +#define SMAASample(tex, coord) tex2D(tex, coord) +#define SMAASamplePoint(tex, coord) tex2D(tex, coord) +#define SMAASampleOffset(tex, coord, offset) tex2D(tex, coord + offset * SMAA_RT_METRICS.xy) +#define SMAA_FLATTEN [flatten] +#define SMAA_BRANCH [branch] +#endif +#if defined(SMAA_HLSL_4) || defined(SMAA_HLSL_4_1) +SamplerState LinearSampler { Filter = MIN_MAG_LINEAR_MIP_POINT; AddressU = Clamp; AddressV = Clamp; }; +SamplerState PointSampler { Filter = MIN_MAG_MIP_POINT; AddressU = Clamp; AddressV = Clamp; }; +#define SMAATexture2D(tex) Texture2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) tex.SampleLevel(LinearSampler, coord, 0) +#define SMAASampleLevelZeroPoint(tex, coord) tex.SampleLevel(PointSampler, coord, 0) +#define SMAASampleLevelZeroOffset(tex, coord, offset) tex.SampleLevel(LinearSampler, coord, 0, offset) +#define SMAASample(tex, coord) tex.Sample(LinearSampler, coord) +#define SMAASamplePoint(tex, coord) tex.Sample(PointSampler, coord) +#define SMAASampleOffset(tex, coord, offset) tex.Sample(LinearSampler, coord, offset) +#define SMAA_FLATTEN [flatten] +#define SMAA_BRANCH [branch] +#define SMAATexture2DMS2(tex) Texture2DMS tex +#define SMAALoad(tex, pos, sample) tex.Load(pos, sample) +#if defined(SMAA_HLSL_4_1) +#define SMAAGather(tex, coord) tex.Gather(LinearSampler, coord, 0) +#endif +#endif +#if defined(SMAA_GLSL_3) || defined(SMAA_GLSL_4) +#define SMAATexture2D(tex) sampler2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) textureLod(tex, coord, 0.0) +#define SMAASampleLevelZeroPoint(tex, coord) textureLod(tex, coord, 0.0) +#define SMAASampleLevelZeroOffset(tex, coord, offset) textureLodOffset(tex, coord, 0.0, offset) +#define SMAASample(tex, coord) texture(tex, coord) +#define SMAASamplePoint(tex, coord) texture(tex, coord) +#define SMAASampleOffset(tex, coord, offset) texture(tex, coord, offset) +#define SMAA_FLATTEN +#define SMAA_BRANCH +#define lerp(a, b, t) mix(a, b, t) +#define saturate(a) clamp(a, 0.0, 1.0) +#if defined(SMAA_GLSL_4) +#define mad(a, b, c) fma(a, b, c) +#define SMAAGather(tex, coord) textureGather(tex, coord) +#else +#define mad(a, b, c) (a * b + c) +#endif +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define int2 ivec2 +#define int3 ivec3 +#define int4 ivec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#endif + +#if !defined(SMAA_HLSL_3) && !defined(SMAA_HLSL_4) && !defined(SMAA_HLSL_4_1) && !defined(SMAA_GLSL_3) && !defined(SMAA_GLSL_4) && !defined(SMAA_CUSTOM_SL) +#error you must define the shading language: SMAA_HLSL_*, SMAA_GLSL_* or SMAA_CUSTOM_SL +#endif + +//----------------------------------------------------------------------------- +// Misc functions + +/** + * Gathers current pixel, and the top-left neighbors. + */ +float3 SMAAGatherNeighbours(float2 texcoord, + float4 offset[3], + SMAATexture2D(tex)) { + #ifdef SMAAGather + return SMAAGather(tex, texcoord + SMAA_RT_METRICS.xy * float2(-0.5, -0.5)).grb; + #else + float P = SMAASamplePoint(tex, texcoord).r; + float Pleft = SMAASamplePoint(tex, offset[0].xy).r; + float Ptop = SMAASamplePoint(tex, offset[0].zw).r; + return float3(P, Pleft, Ptop); + #endif +} + +/** + * Adjusts the threshold by means of predication. + */ +float2 SMAACalculatePredicatedThreshold(float2 texcoord, + float4 offset[3], + SMAATexture2D(predicationTex)) { + float3 neighbours = SMAAGatherNeighbours(texcoord, offset, SMAATexturePass2D(predicationTex)); + float2 delta = abs(neighbours.xx - neighbours.yz); + float2 edges = step(SMAA_PREDICATION_THRESHOLD, delta); + return SMAA_PREDICATION_SCALE * SMAA_THRESHOLD * (1.0 - SMAA_PREDICATION_STRENGTH * edges); +} + +/** + * Conditional move: + */ +void SMAAMovc(bool2 cond, inout float2 variable, float2 value) { + SMAA_FLATTEN if (cond.x) variable.x = value.x; + SMAA_FLATTEN if (cond.y) variable.y = value.y; +} + +void SMAAMovc(bool4 cond, inout float4 variable, float4 value) { + SMAAMovc(cond.xy, variable.xy, value.xy); + SMAAMovc(cond.zw, variable.zw, value.zw); +} + + +#if SMAA_INCLUDE_VS +//----------------------------------------------------------------------------- +// Vertex Shaders + +/** + * Edge Detection Vertex Shader + */ +void SMAAEdgeDetectionVS(float2 texcoord, + out float4 offset[3]) { + offset[0] = mad(SMAA_RT_METRICS.xyxy, float4(-1.0, 0.0, 0.0, -1.0), texcoord.xyxy); + offset[1] = mad(SMAA_RT_METRICS.xyxy, float4( 1.0, 0.0, 0.0, 1.0), texcoord.xyxy); + offset[2] = mad(SMAA_RT_METRICS.xyxy, float4(-2.0, 0.0, 0.0, -2.0), texcoord.xyxy); +} + +/** + * Blend Weight Calculation Vertex Shader + */ +void SMAABlendingWeightCalculationVS(float2 texcoord, + out float2 pixcoord, + out float4 offset[3]) { + pixcoord = texcoord * SMAA_RT_METRICS.zw; + + // We will use these offsets for the searches later on (see @PSEUDO_GATHER4): + offset[0] = mad(SMAA_RT_METRICS.xyxy, float4(-0.25, -0.125, 1.25, -0.125), texcoord.xyxy); + offset[1] = mad(SMAA_RT_METRICS.xyxy, float4(-0.125, -0.25, -0.125, 1.25), texcoord.xyxy); + + // And these for the searches, they indicate the ends of the loops: + offset[2] = mad(SMAA_RT_METRICS.xxyy, + float4(-2.0, 2.0, -2.0, 2.0) * float(SMAA_MAX_SEARCH_STEPS), + float4(offset[0].xz, offset[1].yw)); +} + +/** + * Neighborhood Blending Vertex Shader + */ +void SMAANeighborhoodBlendingVS(float2 texcoord, + out float4 offset) { + offset = mad(SMAA_RT_METRICS.xyxy, float4( 1.0, 0.0, 0.0, 1.0), texcoord.xyxy); +} +#endif // SMAA_INCLUDE_VS + +#if SMAA_INCLUDE_PS +//----------------------------------------------------------------------------- +// Edge Detection Pixel Shaders (First Pass) + +/** + * Luma Edge Detection + * + * IMPORTANT NOTICE: luma edge detection requires gamma-corrected colors, and + * thus 'colorTex' should be a non-sRGB texture. + */ +float2 SMAALumaEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(colorTex) + #if SMAA_PREDICATION + , SMAATexture2D(predicationTex) + #endif + ) { + // Calculate the threshold: + #if SMAA_PREDICATION + float2 threshold = SMAACalculatePredicatedThreshold(texcoord, offset, SMAATexturePass2D(predicationTex)); + #else + float2 threshold = float2(SMAA_THRESHOLD, SMAA_THRESHOLD); + #endif + + // Calculate lumas: + float3 weights = float3(0.2126, 0.7152, 0.0722); + float L = dot(SMAASamplePoint(colorTex, texcoord).rgb, weights); + + float Lleft = dot(SMAASamplePoint(colorTex, offset[0].xy).rgb, weights); + float Ltop = dot(SMAASamplePoint(colorTex, offset[0].zw).rgb, weights); + + // We do the usual threshold: + float4 delta; + delta.xy = abs(L - float2(Lleft, Ltop)); + float2 edges = step(threshold, delta.xy); + + // Then discard if there is no edge: + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + // Calculate right and bottom deltas: + float Lright = dot(SMAASamplePoint(colorTex, offset[1].xy).rgb, weights); + float Lbottom = dot(SMAASamplePoint(colorTex, offset[1].zw).rgb, weights); + delta.zw = abs(L - float2(Lright, Lbottom)); + + // Calculate the maximum delta in the direct neighborhood: + float2 maxDelta = max(delta.xy, delta.zw); + + // Calculate left-left and top-top deltas: + float Lleftleft = dot(SMAASamplePoint(colorTex, offset[2].xy).rgb, weights); + float Ltoptop = dot(SMAASamplePoint(colorTex, offset[2].zw).rgb, weights); + delta.zw = abs(float2(Lleft, Ltop) - float2(Lleftleft, Ltoptop)); + + // Calculate the final maximum delta: + maxDelta = max(maxDelta.xy, delta.zw); + float finalDelta = max(maxDelta.x, maxDelta.y); + + // Local contrast adaptation: + edges.xy *= step(finalDelta, SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR * delta.xy); + + return edges; +} + +/** + * Color Edge Detection + * + * IMPORTANT NOTICE: color edge detection requires gamma-corrected colors, and + * thus 'colorTex' should be a non-sRGB texture. + */ +float2 SMAAColorEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(colorTex) + #if SMAA_PREDICATION + , SMAATexture2D(predicationTex) + #endif + ) { + // Calculate the threshold: + #if SMAA_PREDICATION + float2 threshold = SMAACalculatePredicatedThreshold(texcoord, offset, predicationTex); + #else + float2 threshold = float2(SMAA_THRESHOLD, SMAA_THRESHOLD); + #endif + + // Calculate color deltas: + float4 delta; + float3 C = SMAASamplePoint(colorTex, texcoord).rgb; + + float3 Cleft = SMAASamplePoint(colorTex, offset[0].xy).rgb; + float3 t = abs(C - Cleft); + delta.x = max(max(t.r, t.g), t.b); + + float3 Ctop = SMAASamplePoint(colorTex, offset[0].zw).rgb; + t = abs(C - Ctop); + delta.y = max(max(t.r, t.g), t.b); + + // We do the usual threshold: + float2 edges = step(threshold, delta.xy); + + // Then discard if there is no edge: + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + // Calculate right and bottom deltas: + float3 Cright = SMAASamplePoint(colorTex, offset[1].xy).rgb; + t = abs(C - Cright); + delta.z = max(max(t.r, t.g), t.b); + + float3 Cbottom = SMAASamplePoint(colorTex, offset[1].zw).rgb; + t = abs(C - Cbottom); + delta.w = max(max(t.r, t.g), t.b); + + // Calculate the maximum delta in the direct neighborhood: + float2 maxDelta = max(delta.xy, delta.zw); + + // Calculate left-left and top-top deltas: + float3 Cleftleft = SMAASamplePoint(colorTex, offset[2].xy).rgb; + t = abs(C - Cleftleft); + delta.z = max(max(t.r, t.g), t.b); + + float3 Ctoptop = SMAASamplePoint(colorTex, offset[2].zw).rgb; + t = abs(C - Ctoptop); + delta.w = max(max(t.r, t.g), t.b); + + // Calculate the final maximum delta: + maxDelta = max(maxDelta.xy, delta.zw); + float finalDelta = max(maxDelta.x, maxDelta.y); + + // Local contrast adaptation: + edges.xy *= step(finalDelta, SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR * delta.xy); + + return edges; +} + +/** + * Depth Edge Detection + */ +float2 SMAADepthEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(depthTex)) { + float3 neighbours = SMAAGatherNeighbours(texcoord, offset, SMAATexturePass2D(depthTex)); + float2 delta = abs(neighbours.xx - float2(neighbours.y, neighbours.z)); + float2 edges = step(SMAA_DEPTH_THRESHOLD, delta); + + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + return edges; +} + +//----------------------------------------------------------------------------- +// Diagonal Search Functions + +#if !defined(SMAA_DISABLE_DIAG_DETECTION) + +/** + * Allows to decode two binary values from a bilinear-filtered access. + */ +float2 SMAADecodeDiagBilinearAccess(float2 e) { + // Bilinear access for fetching 'e' have a 0.25 offset, and we are + // interested in the R and G edges: + // + // +---G---+-------+ + // | x o R x | + // +-------+-------+ + // + // Then, if one of these edge is enabled: + // Red: (0.75 * X + 0.25 * 1) => 0.25 or 1.0 + // Green: (0.75 * 1 + 0.25 * X) => 0.75 or 1.0 + // + // This function will unpack the values (mad + mul + round): + // wolframalpha.com: round(x * abs(5 * x - 5 * 0.75)) plot 0 to 1 + e.r = e.r * abs(5.0 * e.r - 5.0 * 0.75); + return round(e); +} + +float4 SMAADecodeDiagBilinearAccess(float4 e) { + e.rb = e.rb * abs(5.0 * e.rb - 5.0 * 0.75); + return round(e); +} + +/** + * These functions allows to perform diagonal pattern searches. + */ +float2 SMAASearchDiag1(SMAATexture2D(edgesTex), float2 texcoord, float2 dir, out float2 e) { + float4 coord = float4(texcoord, -1.0, 1.0); + float3 t = float3(SMAA_RT_METRICS.xy, 1.0); + while (coord.z < float(SMAA_MAX_SEARCH_STEPS_DIAG - 1) && + coord.w > 0.9) { + coord.xyz = mad(t, float3(dir, 1.0), coord.xyz); + e = SMAASampleLevelZero(edgesTex, coord.xy).rg; + coord.w = dot(e, float2(0.5, 0.5)); + } + return coord.zw; +} + +float2 SMAASearchDiag2(SMAATexture2D(edgesTex), float2 texcoord, float2 dir, out float2 e) { + float4 coord = float4(texcoord, -1.0, 1.0); + coord.x += 0.25 * SMAA_RT_METRICS.x; // See @SearchDiag2Optimization + float3 t = float3(SMAA_RT_METRICS.xy, 1.0); + while (coord.z < float(SMAA_MAX_SEARCH_STEPS_DIAG - 1) && + coord.w > 0.9) { + coord.xyz = mad(t, float3(dir, 1.0), coord.xyz); + + // @SearchDiag2Optimization + // Fetch both edges at once using bilinear filtering: + e = SMAASampleLevelZero(edgesTex, coord.xy).rg; + e = SMAADecodeDiagBilinearAccess(e); + + // Non-optimized version: + // e.g = SMAASampleLevelZero(edgesTex, coord.xy).g; + // e.r = SMAASampleLevelZeroOffset(edgesTex, coord.xy, int2(1, 0)).r; + + coord.w = dot(e, float2(0.5, 0.5)); + } + return coord.zw; +} + +/** + * Similar to SMAAArea, this calculates the area corresponding to a certain + * diagonal distance and crossing edges 'e'. + */ +float2 SMAAAreaDiag(SMAATexture2D(areaTex), float2 dist, float2 e, float offset) { + float2 texcoord = mad(float2(SMAA_AREATEX_MAX_DISTANCE_DIAG, SMAA_AREATEX_MAX_DISTANCE_DIAG), e, dist); + + // We do a scale and bias for mapping to texel space: + texcoord = mad(SMAA_AREATEX_PIXEL_SIZE, texcoord, 0.5 * SMAA_AREATEX_PIXEL_SIZE); + + // Diagonal areas are on the second half of the texture: + texcoord.x += 0.5; + + // Move to proper place, according to the subpixel offset: + texcoord.y += SMAA_AREATEX_SUBTEX_SIZE * offset; + + // Do it! + return SMAA_AREATEX_SELECT(SMAASampleLevelZero(areaTex, texcoord)); +} + +/** + * This searches for diagonal patterns and returns the corresponding weights. + */ +float2 SMAACalculateDiagWeights(SMAATexture2D(edgesTex), SMAATexture2D(areaTex), float2 texcoord, float2 e, float4 subsampleIndices) { + float2 weights = float2(0.0, 0.0); + + // Search for the line ends: + float4 d; + float2 end; + if (e.r > 0.0) { + d.xz = SMAASearchDiag1(SMAATexturePass2D(edgesTex), texcoord, float2(-1.0, 1.0), end); + d.x += float(end.y > 0.9); + } else + d.xz = float2(0.0, 0.0); + d.yw = SMAASearchDiag1(SMAATexturePass2D(edgesTex), texcoord, float2(1.0, -1.0), end); + + SMAA_BRANCH + if (d.x + d.y > 2.0) { // d.x + d.y + 1 > 3 + // Fetch the crossing edges: + float4 coords = mad(float4(-d.x + 0.25, d.x, d.y, -d.y - 0.25), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + float4 c; + c.xy = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).rg; + c.zw = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).rg; + c.yxwz = SMAADecodeDiagBilinearAccess(c.xyzw); + + // Non-optimized version: + // float4 coords = mad(float4(-d.x, d.x, d.y, -d.y), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + // float4 c; + // c.x = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).g; + // c.y = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2( 0, 0)).r; + // c.z = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).g; + // c.w = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, -1)).r; + + // Merge crossing edges at each side into a single value: + float2 cc = mad(float2(2.0, 2.0), c.xz, c.yw); + + // Remove the crossing edge if we didn't found the end of the line: + SMAAMovc(bool2(step(0.9, d.zw)), cc, float2(0.0, 0.0)); + + // Fetch the areas for this line: + weights += SMAAAreaDiag(SMAATexturePass2D(areaTex), d.xy, cc, subsampleIndices.z); + } + + // Search for the line ends: + d.xz = SMAASearchDiag2(SMAATexturePass2D(edgesTex), texcoord, float2(-1.0, -1.0), end); + if (SMAASampleLevelZeroOffset(edgesTex, texcoord, int2(1, 0)).r > 0.0) { + d.yw = SMAASearchDiag2(SMAATexturePass2D(edgesTex), texcoord, float2(1.0, 1.0), end); + d.y += float(end.y > 0.9); + } else + d.yw = float2(0.0, 0.0); + + SMAA_BRANCH + if (d.x + d.y > 2.0) { // d.x + d.y + 1 > 3 + // Fetch the crossing edges: + float4 coords = mad(float4(-d.x, -d.x, d.y, d.y), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + float4 c; + c.x = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).g; + c.y = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2( 0, -1)).r; + c.zw = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).gr; + float2 cc = mad(float2(2.0, 2.0), c.xz, c.yw); + + // Remove the crossing edge if we didn't found the end of the line: + SMAAMovc(bool2(step(0.9, d.zw)), cc, float2(0.0, 0.0)); + + // Fetch the areas for this line: + weights += SMAAAreaDiag(SMAATexturePass2D(areaTex), d.xy, cc, subsampleIndices.w).gr; + } + + return weights; +} +#endif + +//----------------------------------------------------------------------------- +// Horizontal/Vertical Search Functions + +/** + * This allows to determine how much length should we add in the last step + * of the searches. It takes the bilinearly interpolated edge (see + * @PSEUDO_GATHER4), and adds 0, 1 or 2, depending on which edges and + * crossing edges are active. + */ +float SMAASearchLength(SMAATexture2D(searchTex), float2 e, float offset) { + // The texture is flipped vertically, with left and right cases taking half + // of the space horizontally: + float2 scale = SMAA_SEARCHTEX_SIZE * float2(0.5, -1.0); + float2 bias = SMAA_SEARCHTEX_SIZE * float2(offset, 1.0); + + // Scale and bias to access texel centers: + scale += float2(-1.0, 1.0); + bias += float2( 0.5, -0.5); + + // Convert from pixel coordinates to texcoords: + // (We use SMAA_SEARCHTEX_PACKED_SIZE because the texture is cropped) + scale *= 1.0 / SMAA_SEARCHTEX_PACKED_SIZE; + bias *= 1.0 / SMAA_SEARCHTEX_PACKED_SIZE; + + // Lookup the search texture: + return SMAA_SEARCHTEX_SELECT(SMAASampleLevelZero(searchTex, mad(scale, e, bias))); +} + +/** + * Horizontal/vertical search functions for the 2nd pass. + */ +float SMAASearchXLeft(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + /** + * @PSEUDO_GATHER4 + * This texcoord has been offset by (-0.25, -0.125) in the vertex shader to + * sample between edge, thus fetching four edges in a row. + * Sampling with different offsets in each direction allows to disambiguate + * which edges are active from the four fetched ones. + */ + float2 e = float2(0.0, 1.0); + while (texcoord.x > end && + e.g > 0.8281 && // Is there some edge not activated? + e.r == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(-float2(2.0, 0.0), SMAA_RT_METRICS.xy, texcoord); + } + + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.0), 3.25); + return mad(SMAA_RT_METRICS.x, offset, texcoord.x); + + // Non-optimized version: + // We correct the previous (-0.25, -0.125) offset we applied: + // texcoord.x += 0.25 * SMAA_RT_METRICS.x; + + // The searches are bias by 1, so adjust the coords accordingly: + // texcoord.x += SMAA_RT_METRICS.x; + + // Disambiguate the length added by the last step: + // texcoord.x += 2.0 * SMAA_RT_METRICS.x; // Undo last step + // texcoord.x -= SMAA_RT_METRICS.x * (255.0 / 127.0) * SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.0); + // return mad(SMAA_RT_METRICS.x, offset, texcoord.x); +} + +float SMAASearchXRight(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(0.0, 1.0); + while (texcoord.x < end && + e.g > 0.8281 && // Is there some edge not activated? + e.r == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(float2(2.0, 0.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.5), 3.25); + return mad(-SMAA_RT_METRICS.x, offset, texcoord.x); +} + +float SMAASearchYUp(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(1.0, 0.0); + while (texcoord.y > end && + e.r > 0.8281 && // Is there some edge not activated? + e.g == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(-float2(0.0, 2.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e.gr, 0.0), 3.25); + return mad(SMAA_RT_METRICS.y, offset, texcoord.y); +} + +float SMAASearchYDown(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(1.0, 0.0); + while (texcoord.y < end && + e.r > 0.8281 && // Is there some edge not activated? + e.g == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(float2(0.0, 2.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e.gr, 0.5), 3.25); + return mad(-SMAA_RT_METRICS.y, offset, texcoord.y); +} + +/** + * Ok, we have the distance and both crossing edges. So, what are the areas + * at each side of current edge? + */ +float2 SMAAArea(SMAATexture2D(areaTex), float2 dist, float e1, float e2, float offset) { + // Rounding prevents precision errors of bilinear filtering: + float2 texcoord = mad(float2(SMAA_AREATEX_MAX_DISTANCE, SMAA_AREATEX_MAX_DISTANCE), round(4.0 * float2(e1, e2)), dist); + + // We do a scale and bias for mapping to texel space: + texcoord = mad(SMAA_AREATEX_PIXEL_SIZE, texcoord, 0.5 * SMAA_AREATEX_PIXEL_SIZE); + + // Move to proper place, according to the subpixel offset: + texcoord.y = mad(SMAA_AREATEX_SUBTEX_SIZE, offset, texcoord.y); + + // Do it! + return SMAA_AREATEX_SELECT(SMAASampleLevelZero(areaTex, texcoord)); +} + +//----------------------------------------------------------------------------- +// Corner Detection Functions + +void SMAADetectHorizontalCornerPattern(SMAATexture2D(edgesTex), inout float2 weights, float4 texcoord, float2 d) { + #if !defined(SMAA_DISABLE_CORNER_DETECTION) + float2 leftRight = step(d.xy, d.yx); + float2 rounding = (1.0 - SMAA_CORNER_ROUNDING_NORM) * leftRight; + + rounding /= leftRight.x + leftRight.y; // Reduce blending for pixels in the center of a line. + + float2 factor = float2(1.0, 1.0); + factor.x -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(0, 1)).r; + factor.x -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(1, 1)).r; + factor.y -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(0, -2)).r; + factor.y -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(1, -2)).r; + + weights *= saturate(factor); + #endif +} + +void SMAADetectVerticalCornerPattern(SMAATexture2D(edgesTex), inout float2 weights, float4 texcoord, float2 d) { + #if !defined(SMAA_DISABLE_CORNER_DETECTION) + float2 leftRight = step(d.xy, d.yx); + float2 rounding = (1.0 - SMAA_CORNER_ROUNDING_NORM) * leftRight; + + rounding /= leftRight.x + leftRight.y; + + float2 factor = float2(1.0, 1.0); + factor.x -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2( 1, 0)).g; + factor.x -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2( 1, 1)).g; + factor.y -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(-2, 0)).g; + factor.y -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(-2, 1)).g; + + weights *= saturate(factor); + #endif +} + +//----------------------------------------------------------------------------- +// Blending Weight Calculation Pixel Shader (Second Pass) + +float4 SMAABlendingWeightCalculationPS(float2 texcoord, + float2 pixcoord, + float4 offset[3], + SMAATexture2D(edgesTex), + SMAATexture2D(areaTex), + SMAATexture2D(searchTex), + float4 subsampleIndices) { // Just pass zero for SMAA 1x, see @SUBSAMPLE_INDICES. + float4 weights = float4(0.0, 0.0, 0.0, 0.0); + + float2 e = SMAASample(edgesTex, texcoord).rg; + + SMAA_BRANCH + if (e.g > 0.0) { // Edge at north + #if !defined(SMAA_DISABLE_DIAG_DETECTION) + // Diagonals have both north and west edges, so searching for them in + // one of the boundaries is enough. + weights.rg = SMAACalculateDiagWeights(SMAATexturePass2D(edgesTex), SMAATexturePass2D(areaTex), texcoord, e, subsampleIndices); + + // We give priority to diagonals, so if we find a diagonal we skip + // horizontal/vertical processing. + SMAA_BRANCH + if (weights.r == -weights.g) { // weights.r + weights.g == 0.0 + #endif + + float2 d; + + // Find the distance to the left: + float3 coords; + coords.x = SMAASearchXLeft(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[0].xy, offset[2].x); + coords.y = offset[1].y; // offset[1].y = texcoord.y - 0.25 * SMAA_RT_METRICS.y (@CROSSING_OFFSET) + d.x = coords.x; + + // Now fetch the left crossing edges, two at a time using bilinear + // filtering. Sampling at -0.25 (see @CROSSING_OFFSET) enables to + // discern what value each edge has: + float e1 = SMAASampleLevelZero(edgesTex, coords.xy).r; + + // Find the distance to the right: + coords.z = SMAASearchXRight(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[0].zw, offset[2].y); + d.y = coords.z; + + // We want the distances to be in pixel units (doing this here allow to + // better interleave arithmetic and memory accesses): + d = abs(round(mad(SMAA_RT_METRICS.zz, d, -pixcoord.xx))); + + // SMAAArea below needs a sqrt, as the areas texture is compressed + // quadratically: + float2 sqrt_d = sqrt(d); + + // Fetch the right crossing edges: + float e2 = SMAASampleLevelZeroOffset(edgesTex, coords.zy, int2(1, 0)).r; + + // Ok, we know how this pattern looks like, now it is time for getting + // the actual area: + weights.rg = SMAAArea(SMAATexturePass2D(areaTex), sqrt_d, e1, e2, subsampleIndices.y); + + // Fix corners: + coords.y = texcoord.y; + SMAADetectHorizontalCornerPattern(SMAATexturePass2D(edgesTex), weights.rg, coords.xyzy, d); + + #if !defined(SMAA_DISABLE_DIAG_DETECTION) + } else + e.r = 0.0; // Skip vertical processing. + #endif + } + + SMAA_BRANCH + if (e.r > 0.0) { // Edge at west + float2 d; + + // Find the distance to the top: + float3 coords; + coords.y = SMAASearchYUp(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[1].xy, offset[2].z); + coords.x = offset[0].x; // offset[1].x = texcoord.x - 0.25 * SMAA_RT_METRICS.x; + d.x = coords.y; + + // Fetch the top crossing edges: + float e1 = SMAASampleLevelZero(edgesTex, coords.xy).g; + + // Find the distance to the bottom: + coords.z = SMAASearchYDown(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[1].zw, offset[2].w); + d.y = coords.z; + + // We want the distances to be in pixel units: + d = abs(round(mad(SMAA_RT_METRICS.ww, d, -pixcoord.yy))); + + // SMAAArea below needs a sqrt, as the areas texture is compressed + // quadratically: + float2 sqrt_d = sqrt(d); + + // Fetch the bottom crossing edges: + float e2 = SMAASampleLevelZeroOffset(edgesTex, coords.xz, int2(0, 1)).g; + + // Get the area for this direction: + weights.ba = SMAAArea(SMAATexturePass2D(areaTex), sqrt_d, e1, e2, subsampleIndices.x); + + // Fix corners: + coords.x = texcoord.x; + SMAADetectVerticalCornerPattern(SMAATexturePass2D(edgesTex), weights.ba, coords.xyxz, d); + } + + return weights; +} + +//----------------------------------------------------------------------------- +// Neighborhood Blending Pixel Shader (Third Pass) + +float4 SMAANeighborhoodBlendingPS(float2 texcoord, + float4 offset, + SMAATexture2D(colorTex), + SMAATexture2D(blendTex) + #if SMAA_REPROJECTION + , SMAATexture2D(velocityTex) + #endif + ) { + // Fetch the blending weights for current pixel: + float4 a; + a.x = SMAASample(blendTex, offset.xy).a; // Right + a.y = SMAASample(blendTex, offset.zw).g; // Top + a.wz = SMAASample(blendTex, texcoord).xz; // Bottom / Left + + // Is there any blending weight with a value greater than 0.0? + SMAA_BRANCH + if (dot(a, float4(1.0, 1.0, 1.0, 1.0)) < 1e-5) { + float4 color = SMAASampleLevelZero(colorTex, texcoord); + + #if SMAA_REPROJECTION + float2 velocity = SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, texcoord)); + + // Pack velocity into the alpha channel: + color.a = sqrt(5.0 * length(velocity)); + #endif + + return color; + } else { + bool h = max(a.x, a.z) > max(a.y, a.w); // max(horizontal) > max(vertical) + + // Calculate the blending offsets: + float4 blendingOffset = float4(0.0, a.y, 0.0, a.w); + float2 blendingWeight = a.yw; + SMAAMovc(bool4(h, h, h, h), blendingOffset, float4(a.x, 0.0, a.z, 0.0)); + SMAAMovc(bool2(h, h), blendingWeight, a.xz); + blendingWeight /= dot(blendingWeight, float2(1.0, 1.0)); + + // Calculate the texture coordinates: + float4 blendingCoord = mad(blendingOffset, float4(SMAA_RT_METRICS.xy, -SMAA_RT_METRICS.xy), texcoord.xyxy); + + // We exploit bilinear filtering to mix current pixel with the chosen + // neighbor: + float4 color = blendingWeight.x * SMAASampleLevelZero(colorTex, blendingCoord.xy); + color += blendingWeight.y * SMAASampleLevelZero(colorTex, blendingCoord.zw); + + #if SMAA_REPROJECTION + // Antialias velocity for proper reprojection in a later stage: + float2 velocity = blendingWeight.x * SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, blendingCoord.xy)); + velocity += blendingWeight.y * SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, blendingCoord.zw)); + + // Pack velocity into the alpha channel: + color.a = sqrt(5.0 * length(velocity)); + #endif + + return color; + } +} + +//----------------------------------------------------------------------------- +// Temporal Resolve Pixel Shader (Optional Pass) + +float4 SMAAResolvePS(float2 texcoord, + SMAATexture2D(currentColorTex), + SMAATexture2D(previousColorTex) + #if SMAA_REPROJECTION + , SMAATexture2D(velocityTex) + #endif + ) { + #if SMAA_REPROJECTION + // Velocity is assumed to be calculated for motion blur, so we need to + // inverse it for reprojection: + float2 velocity = -SMAA_DECODE_VELOCITY(SMAASamplePoint(velocityTex, texcoord).rg); + + // Fetch current pixel: + float4 current = SMAASamplePoint(currentColorTex, texcoord); + + // Reproject current coordinates and fetch previous pixel: + float4 previous = SMAASamplePoint(previousColorTex, texcoord + velocity); + + // Attenuate the previous pixel if the velocity is different: + float delta = abs(current.a * current.a - previous.a * previous.a) / 5.0; + float weight = 0.5 * saturate(1.0 - sqrt(delta) * SMAA_REPROJECTION_WEIGHT_SCALE); + + // Blend the pixels according to the calculated weight: + return lerp(current, previous, weight); + #else + // Just blend the pixels: + float4 current = SMAASamplePoint(currentColorTex, texcoord); + float4 previous = SMAASamplePoint(previousColorTex, texcoord); + return lerp(current, previous, 0.5); + #endif +} + +//----------------------------------------------------------------------------- +// Separate Multisamples Pixel Shader (Optional Pass) + +#ifdef SMAALoad +void SMAASeparatePS(float4 position, + float2 texcoord, + out float4 target0, + out float4 target1, + SMAATexture2DMS2(colorTexMS)) { + int2 pos = int2(position.xy); + target0 = SMAALoad(colorTexMS, pos, 0); + target1 = SMAALoad(colorTexMS, pos, 1); +} +#endif + +//----------------------------------------------------------------------------- +#endif // SMAA_INCLUDE_PS diff --git a/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_blend.glsl b/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_blend.glsl new file mode 100644 index 00000000..c875ce12 --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_blend.glsl @@ -0,0 +1,26 @@ +layout(rgba8, binding = 0) uniform image2D imgOutput; + +uniform sampler2D inputTexture; +layout( location=0 ) uniform vec2 invResolution; +uniform sampler2D samplerArea; +uniform sampler2D samplerSearch; + +void main() { + ivec2 loc = ivec2(gl_GlobalInvocationID.x * 4, gl_GlobalInvocationID.y * 4); + for(int i = 0; i < 4; i++) + { + for(int j = 0; j < 4; j++) + { + ivec2 texelCoord = ivec2(loc.x + i, loc.y + j); + vec2 coord = (texelCoord + vec2(0.5)) / invResolution; + vec2 pixCoord; + vec4 offset[3]; + + SMAABlendingWeightCalculationVS(coord, pixCoord, offset); + + vec4 oColor = SMAABlendingWeightCalculationPS(coord, pixCoord, offset, inputTexture, samplerArea, samplerSearch, ivec4(0)); + + imageStore(imgOutput, texelCoord, oColor); + } + } +} diff --git a/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_edge.glsl b/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_edge.glsl new file mode 100644 index 00000000..fd5d9715 --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_edge.glsl @@ -0,0 +1,24 @@ +layout(rgba8, binding = 0) uniform image2D imgOutput; + +uniform sampler2D inputTexture; +layout( location=0 ) uniform vec2 invResolution; + +void main() +{ + vec2 loc = ivec2(gl_GlobalInvocationID.x * 4, gl_GlobalInvocationID.y * 4); + for(int i = 0; i < 4; i++) + { + for(int j = 0; j < 4; j++) + { + ivec2 texelCoord = ivec2(loc.x + i, loc.y + j); + vec2 coord = (texelCoord + vec2(0.5)) / invResolution; + vec4 offset[3]; + SMAAEdgeDetectionVS(coord, offset); + vec2 oColor = SMAAColorEdgeDetectionPS(coord, offset, inputTexture); + if (oColor != float2(-2.0, -2.0)) + { + imageStore(imgOutput, texelCoord, vec4(oColor, 0.0, 1.0)); + } + } + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_neighbour.glsl b/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_neighbour.glsl new file mode 100644 index 00000000..2e9432ae --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_neighbour.glsl @@ -0,0 +1,26 @@ +layout(rgba8, binding = 0) uniform image2D imgOutput; + +uniform sampler2D inputTexture; +layout( location=0 ) uniform vec2 invResolution; +uniform sampler2D samplerBlend; + +void main() { + vec2 loc = ivec2(gl_GlobalInvocationID.x * 4, gl_GlobalInvocationID.y * 4); + for(int i = 0; i < 4; i++) + { + for(int j = 0; j < 4; j++) + { + ivec2 texelCoord = ivec2(loc.x + i, loc.y + j); + vec2 coord = (texelCoord + vec2(0.5)) / invResolution; + vec2 pixCoord; + vec4 offset; + + SMAANeighborhoodBlendingVS(coord, offset); + + vec4 oColor = SMAANeighborhoodBlendingPS(coord, offset, inputTexture, samplerBlend); + + imageStore(imgOutput, texelCoord, oColor); + } + } + +} diff --git a/Ryujinx.Graphics.OpenGL/Effects/SmaaPostProcessingEffect.cs b/Ryujinx.Graphics.OpenGL/Effects/SmaaPostProcessingEffect.cs new file mode 100644 index 00000000..1ad300c8 --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Effects/SmaaPostProcessingEffect.cs @@ -0,0 +1,261 @@ +using OpenTK.Graphics.OpenGL; +using Ryujinx.Common; +using Ryujinx.Graphics.GAL; +using Ryujinx.Graphics.OpenGL.Image; +using System; + +namespace Ryujinx.Graphics.OpenGL.Effects.Smaa +{ + internal partial class SmaaPostProcessingEffect : IPostProcessingEffect + { + public const int AreaWidth = 160; + public const int AreaHeight = 560; + public const int SearchWidth = 64; + public const int SearchHeight = 16; + + private readonly OpenGLRenderer _renderer; + private TextureStorage _outputTexture; + private TextureStorage _searchTexture; + private TextureStorage _areaTexture; + private int[] _edgeShaderPrograms; + private int[] _blendShaderPrograms; + private int[] _neighbourShaderPrograms; + private TextureStorage _edgeOutputTexture; + private TextureStorage _blendOutputTexture; + private string[] _qualities; + private int _inputUniform; + private int _outputUniform; + private int _samplerAreaUniform; + private int _samplerSearchUniform; + private int _samplerBlendUniform; + private int _resolutionUniform; + private int _quality = 1; + + public int Quality + { + get => _quality; set + { + _quality = Math.Clamp(value, 0, _qualities.Length - 1); + } + } + public SmaaPostProcessingEffect(OpenGLRenderer renderer, int quality) + { + _renderer = renderer; + + _edgeShaderPrograms = Array.Empty(); + _blendShaderPrograms = Array.Empty(); + _neighbourShaderPrograms = Array.Empty(); + + _qualities = new string[] { "SMAA_PRESET_LOW", "SMAA_PRESET_MEDIUM", "SMAA_PRESET_HIGH", "SMAA_PRESET_ULTRA" }; + + Quality = quality; + + Initialize(); + } + + public void Dispose() + { + _searchTexture?.Dispose(); + _areaTexture?.Dispose(); + _outputTexture?.Dispose(); + _edgeOutputTexture?.Dispose(); + _blendOutputTexture?.Dispose(); + + DeleteShaders(); + } + + private void DeleteShaders() + { + for (int i = 0; i < _edgeShaderPrograms.Length; i++) + { + GL.DeleteProgram(_edgeShaderPrograms[i]); + GL.DeleteProgram(_blendShaderPrograms[i]); + GL.DeleteProgram(_neighbourShaderPrograms[i]); + } + } + + private unsafe void RecreateShaders(int width, int height) + { + string baseShader = EmbeddedResources.ReadAllText("Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa.hlsl"); + var pixelSizeDefine = $"#define SMAA_RT_METRICS float4(1.0 / {width}.0, 1.0 / {height}.0, {width}, {height}) \n"; + + _edgeShaderPrograms = new int[_qualities.Length]; + _blendShaderPrograms = new int[_qualities.Length]; + _neighbourShaderPrograms = new int[_qualities.Length]; + + for (int i = 0; i < +_edgeShaderPrograms.Length; i++) + { + var presets = $"#version 430 core \n#define {_qualities[i]} 1 \n{pixelSizeDefine}#define SMAA_GLSL_4 1 \nlayout (local_size_x = 16, local_size_y = 16) in;\n{baseShader}"; + + var edgeShaderData = EmbeddedResources.ReadAllText("Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_edge.glsl"); + var blendShaderData = EmbeddedResources.ReadAllText("Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_blend.glsl"); + var neighbourShaderData = EmbeddedResources.ReadAllText("Ryujinx.Graphics.OpenGL/Effects/Shaders/smaa_neighbour.glsl"); + + var shaders = new string[] { presets, edgeShaderData }; + var edgeProgram = ShaderHelper.CompileProgram(shaders, ShaderType.ComputeShader); + + shaders[1] = blendShaderData; + var blendProgram = ShaderHelper.CompileProgram(shaders, ShaderType.ComputeShader); + + shaders[1] = neighbourShaderData; + var neighbourProgram = ShaderHelper.CompileProgram(shaders, ShaderType.ComputeShader); + + _edgeShaderPrograms[i] = edgeProgram; + _blendShaderPrograms[i] = blendProgram; + _neighbourShaderPrograms[i] = neighbourProgram; + } + + _inputUniform = GL.GetUniformLocation(_edgeShaderPrograms[0], "inputTexture"); + _outputUniform = GL.GetUniformLocation(_edgeShaderPrograms[0], "imgOutput"); + _samplerAreaUniform = GL.GetUniformLocation(_blendShaderPrograms[0], "samplerArea"); + _samplerSearchUniform = GL.GetUniformLocation(_blendShaderPrograms[0], "samplerSearch"); + _samplerBlendUniform = GL.GetUniformLocation(_neighbourShaderPrograms[0], "samplerBlend"); + _resolutionUniform = GL.GetUniformLocation(_edgeShaderPrograms[0], "invResolution"); + } + + private void Initialize() + { + var areaInfo = new TextureCreateInfo(AreaWidth, + AreaHeight, + 1, + 1, + 1, + 1, + 1, + 1, + Format.R8G8Unorm, + DepthStencilMode.Depth, + Target.Texture2D, + SwizzleComponent.Red, + SwizzleComponent.Green, + SwizzleComponent.Blue, + SwizzleComponent.Alpha); + + var searchInfo = new TextureCreateInfo(SearchWidth, + SearchHeight, + 1, + 1, + 1, + 1, + 1, + 1, + Format.R8Unorm, + DepthStencilMode.Depth, + Target.Texture2D, + SwizzleComponent.Red, + SwizzleComponent.Green, + SwizzleComponent.Blue, + SwizzleComponent.Alpha); + + _areaTexture = new TextureStorage(_renderer, areaInfo, 1); + _searchTexture = new TextureStorage(_renderer, searchInfo, 1); + + var areaTexture = EmbeddedResources.Read("Ryujinx.Graphics.OpenGL/Effects/Textures/SmaaAreaTexture.bin"); + var searchTexture = EmbeddedResources.Read("Ryujinx.Graphics.OpenGL/Effects/Textures/SmaaSearchTexture.bin"); + + var areaView = _areaTexture.CreateDefaultView(); + var searchView = _searchTexture.CreateDefaultView(); + + areaView.SetData(areaTexture); + searchView.SetData(searchTexture); + } + + public TextureView Run(TextureView view, int width, int height) + { + if (_outputTexture == null || _outputTexture.Info.Width != view.Width || _outputTexture.Info.Height != view.Height) + { + _outputTexture?.Dispose(); + _outputTexture = new TextureStorage(_renderer, view.Info, view.ScaleFactor); + _outputTexture.CreateDefaultView(); + _edgeOutputTexture = new TextureStorage(_renderer, view.Info, view.ScaleFactor); + _edgeOutputTexture.CreateDefaultView(); + _blendOutputTexture = new TextureStorage(_renderer, view.Info, view.ScaleFactor); + _blendOutputTexture.CreateDefaultView(); + + DeleteShaders(); + + RecreateShaders(view.Width, view.Height); + } + + var textureView = _outputTexture.CreateView(view.Info, 0, 0) as TextureView; + var edgeOutput = _edgeOutputTexture.DefaultView as TextureView; + var blendOutput = _blendOutputTexture.DefaultView as TextureView; + var areaTexture = _areaTexture.DefaultView as TextureView; + var searchTexture = _searchTexture.DefaultView as TextureView; + + var previousFramebuffer = GL.GetInteger(GetPName.FramebufferBinding); + int previousUnit = GL.GetInteger(GetPName.ActiveTexture); + GL.ActiveTexture(TextureUnit.Texture0); + int previousTextureBinding0 = GL.GetInteger(GetPName.TextureBinding2D); + GL.ActiveTexture(TextureUnit.Texture1); + int previousTextureBinding1 = GL.GetInteger(GetPName.TextureBinding2D); + GL.ActiveTexture(TextureUnit.Texture2); + int previousTextureBinding2 = GL.GetInteger(GetPName.TextureBinding2D); + + var framebuffer = new Framebuffer(); + framebuffer.Bind(); + framebuffer.AttachColor(0, edgeOutput); + GL.Clear(ClearBufferMask.ColorBufferBit); + GL.ClearColor(0, 0, 0, 0); + framebuffer.AttachColor(0, blendOutput); + GL.Clear(ClearBufferMask.ColorBufferBit); + GL.ClearColor(0, 0, 0, 0); + + GL.BindFramebuffer(FramebufferTarget.Framebuffer, previousFramebuffer); + + framebuffer.Dispose(); + + var dispatchX = BitUtils.DivRoundUp(view.Width, IPostProcessingEffect.LocalGroupSize); + var dispatchY = BitUtils.DivRoundUp(view.Height, IPostProcessingEffect.LocalGroupSize); + + int previousProgram = GL.GetInteger(GetPName.CurrentProgram); + GL.BindImageTexture(0, edgeOutput.Handle, 0, false, 0, TextureAccess.ReadWrite, SizedInternalFormat.Rgba8); + GL.UseProgram(_edgeShaderPrograms[Quality]); + view.Bind(0); + GL.Uniform1(_inputUniform, 0); + GL.Uniform1(_outputUniform, 0); + GL.Uniform2(_resolutionUniform, (float)view.Width, (float)view.Height); + GL.DispatchCompute(dispatchX, dispatchY, 1); + GL.MemoryBarrier(MemoryBarrierFlags.ShaderImageAccessBarrierBit); + + GL.BindImageTexture(0, blendOutput.Handle, 0, false, 0, TextureAccess.ReadWrite, SizedInternalFormat.Rgba8); + GL.UseProgram(_blendShaderPrograms[Quality]); + edgeOutput.Bind(0); + areaTexture.Bind(1); + searchTexture.Bind(2); + GL.Uniform1(_inputUniform, 0); + GL.Uniform1(_outputUniform, 0); + GL.Uniform1(_samplerAreaUniform, 1); + GL.Uniform1(_samplerSearchUniform, 2); + GL.Uniform2(_resolutionUniform, (float)view.Width, (float)view.Height); + GL.DispatchCompute(dispatchX, dispatchY, 1); + GL.MemoryBarrier(MemoryBarrierFlags.ShaderImageAccessBarrierBit); + + GL.BindImageTexture(0, textureView.Handle, 0, false, 0, TextureAccess.ReadWrite, SizedInternalFormat.Rgba8); + GL.UseProgram(_neighbourShaderPrograms[Quality]); + view.Bind(0); + blendOutput.Bind(1); + GL.Uniform1(_inputUniform, 0); + GL.Uniform1(_outputUniform, 0); + GL.Uniform1(_samplerBlendUniform, 1); + GL.Uniform2(_resolutionUniform, (float)view.Width, (float)view.Height); + GL.DispatchCompute(dispatchX, dispatchY, 1); + GL.MemoryBarrier(MemoryBarrierFlags.ShaderImageAccessBarrierBit); + + (_renderer.Pipeline as Pipeline).RestoreImages1And2(); + + GL.UseProgram(previousProgram); + + GL.ActiveTexture(TextureUnit.Texture0); + GL.BindTexture(TextureTarget.Texture2D, previousTextureBinding0); + GL.ActiveTexture(TextureUnit.Texture1); + GL.BindTexture(TextureTarget.Texture2D, previousTextureBinding1); + GL.ActiveTexture(TextureUnit.Texture2); + GL.BindTexture(TextureTarget.Texture2D, previousTextureBinding2); + + GL.ActiveTexture((TextureUnit)previousUnit); + + return textureView; + } + } +} diff --git a/Ryujinx.Graphics.OpenGL/Effects/Textures/SmaaAreaTexture.bin b/Ryujinx.Graphics.OpenGL/Effects/Textures/SmaaAreaTexture.bin new file mode 100644 index 00000000..f4a7a1b4 Binary files /dev/null and b/Ryujinx.Graphics.OpenGL/Effects/Textures/SmaaAreaTexture.bin differ diff --git a/Ryujinx.Graphics.OpenGL/Effects/Textures/SmaaSearchTexture.bin b/Ryujinx.Graphics.OpenGL/Effects/Textures/SmaaSearchTexture.bin new file mode 100644 index 00000000..db5bf73f Binary files /dev/null and b/Ryujinx.Graphics.OpenGL/Effects/Textures/SmaaSearchTexture.bin differ diff --git a/Ryujinx.Graphics.OpenGL/Ryujinx.Graphics.OpenGL.csproj b/Ryujinx.Graphics.OpenGL/Ryujinx.Graphics.OpenGL.csproj index 9fd2c48a..2313cc68 100644 --- a/Ryujinx.Graphics.OpenGL/Ryujinx.Graphics.OpenGL.csproj +++ b/Ryujinx.Graphics.OpenGL/Ryujinx.Graphics.OpenGL.csproj @@ -9,6 +9,20 @@ + + + + + + + + + + + + + + diff --git a/Ryujinx.Graphics.OpenGL/Window.cs b/Ryujinx.Graphics.OpenGL/Window.cs index 8f7917f9..d6606f39 100644 --- a/Ryujinx.Graphics.OpenGL/Window.cs +++ b/Ryujinx.Graphics.OpenGL/Window.cs @@ -1,5 +1,7 @@ using OpenTK.Graphics.OpenGL; using Ryujinx.Graphics.GAL; +using Ryujinx.Graphics.OpenGL.Effects; +using Ryujinx.Graphics.OpenGL.Effects.Smaa; using Ryujinx.Graphics.OpenGL.Image; using System; @@ -7,14 +9,24 @@ namespace Ryujinx.Graphics.OpenGL { class Window : IWindow, IDisposable { - private const int TextureCount = 3; private readonly OpenGLRenderer _renderer; private bool _initialized; private int _width; private int _height; + private bool _updateSize; private int _copyFramebufferHandle; + private IPostProcessingEffect _antiAliasing; + private IScalingFilter _scalingFilter; + private bool _isLinear; + private AntiAliasing _currentAntiAliasing; + private bool _updateEffect; + private ScalingFilter _currentScalingFilter; + private float _scalingFilterLevel; + private bool _updateScalingFilter; + private bool _isBgra; + private TextureView _upscaledTexture; internal BackgroundContextWorker BackgroundContext { get; private set; } @@ -48,6 +60,8 @@ namespace Ryujinx.Graphics.OpenGL { _width = width; _height = height; + + _updateSize = true; } private void CopyTextureToFrameBufferRGB(int drawFramebuffer, int readFramebuffer, TextureView view, ImageCrop crop, Action swapBuffersCallback) @@ -57,6 +71,32 @@ namespace Ryujinx.Graphics.OpenGL TextureView viewConverted = view.Format.IsBgr() ? _renderer.TextureCopy.BgraSwap(view) : view; + UpdateEffect(); + + if (_antiAliasing != null) + { + var oldView = viewConverted; + + viewConverted = _antiAliasing.Run(viewConverted, _width, _height); + + if (viewConverted.Format.IsBgr()) + { + var swappedView = _renderer.TextureCopy.BgraSwap(viewConverted); + + viewConverted?.Dispose(); + + viewConverted = swappedView; + } + + if (viewConverted != oldView && oldView != view) + { + oldView.Dispose(); + } + } + + GL.BindFramebuffer(FramebufferTarget.DrawFramebuffer, drawFramebuffer); + GL.BindFramebuffer(FramebufferTarget.ReadFramebuffer, readFramebuffer); + GL.FramebufferTexture( FramebufferTarget.ReadFramebuffer, FramebufferAttachment.ColorAttachment0, @@ -71,12 +111,12 @@ namespace Ryujinx.Graphics.OpenGL GL.Clear(ClearBufferMask.ColorBufferBit); int srcX0, srcX1, srcY0, srcY1; - float scale = view.ScaleFactor; + float scale = viewConverted.ScaleFactor; if (crop.Left == 0 && crop.Right == 0) { srcX0 = 0; - srcX1 = (int)(view.Width / scale); + srcX1 = (int)(viewConverted.Width / scale); } else { @@ -87,7 +127,7 @@ namespace Ryujinx.Graphics.OpenGL if (crop.Top == 0 && crop.Bottom == 0) { srcY0 = 0; - srcY1 = (int)(view.Height / scale); + srcY1 = (int)(viewConverted.Height / scale); } else { @@ -125,6 +165,42 @@ namespace Ryujinx.Graphics.OpenGL ScreenCaptureRequested = false; } + if (_scalingFilter != null) + { + if (viewConverted.Format.IsBgr() && !_isBgra) + { + RecreateUpscalingTexture(true); + } + + _scalingFilter.Run( + viewConverted, + _upscaledTexture, + _width, + _height, + new Extents2D( + srcX0, + srcY0, + srcX1, + srcY1), + new Extents2D( + dstX0, + dstY0, + dstX1, + dstY1) + ); + + srcX0 = dstX0; + srcY0 = dstY0; + srcX1 = dstX1; + srcY1 = dstY1; + + GL.FramebufferTexture( + FramebufferTarget.ReadFramebuffer, + FramebufferAttachment.ColorAttachment0, + _upscaledTexture.Handle, + 0); + } + GL.BlitFramebuffer( srcX0, srcY0, @@ -135,7 +211,7 @@ namespace Ryujinx.Graphics.OpenGL dstX1, dstY1, ClearBufferMask.ColorBufferBit, - BlitFramebufferFilter.Linear); + _isLinear ? BlitFramebufferFilter.Linear : BlitFramebufferFilter.Nearest); // Remove Alpha channel GL.ColorMask(false, false, false, true); @@ -209,6 +285,135 @@ namespace Ryujinx.Graphics.OpenGL _copyFramebufferHandle = 0; } + + _antiAliasing?.Dispose(); + _scalingFilter?.Dispose(); + _upscaledTexture?.Dispose(); + } + + public void SetAntiAliasing(AntiAliasing effect) + { + if (_currentAntiAliasing == effect && _antiAliasing != null) + { + return; + } + + _currentAntiAliasing = effect; + + _updateEffect = true; + } + + public void SetScalingFilter(ScalingFilter type) + { + if (_currentScalingFilter == type && _antiAliasing != null) + { + return; + } + + _currentScalingFilter = type; + + _updateScalingFilter = true; + } + + private void UpdateEffect() + { + if (_updateEffect) + { + _updateEffect = false; + + switch (_currentAntiAliasing) + { + case AntiAliasing.Fxaa: + _antiAliasing?.Dispose(); + _antiAliasing = new FxaaPostProcessingEffect(_renderer); + break; + case AntiAliasing.None: + _antiAliasing?.Dispose(); + _antiAliasing = null; + break; + case AntiAliasing.SmaaLow: + case AntiAliasing.SmaaMedium: + case AntiAliasing.SmaaHigh: + case AntiAliasing.SmaaUltra: + var quality = _currentAntiAliasing - AntiAliasing.SmaaLow; + if (_antiAliasing is SmaaPostProcessingEffect smaa) + { + smaa.Quality = quality; + } + else + { + _antiAliasing?.Dispose(); + _antiAliasing = new SmaaPostProcessingEffect(_renderer, quality); + } + break; + } + } + + if (_updateSize && !_updateScalingFilter) + { + RecreateUpscalingTexture(); + } + + _updateSize = false; + + if (_updateScalingFilter) + { + _updateScalingFilter = false; + + switch (_currentScalingFilter) + { + case ScalingFilter.Bilinear: + case ScalingFilter.Nearest: + _scalingFilter?.Dispose(); + _scalingFilter = null; + _isLinear = _currentScalingFilter == ScalingFilter.Bilinear; + _upscaledTexture?.Dispose(); + _upscaledTexture = null; + break; + case ScalingFilter.Fsr: + if (_scalingFilter is not FsrScalingFilter) + { + _scalingFilter?.Dispose(); + _scalingFilter = new FsrScalingFilter(_renderer, _antiAliasing); + } + _isLinear = false; + _scalingFilter.Level = _scalingFilterLevel; + + RecreateUpscalingTexture(); + break; + } + } + } + + private void RecreateUpscalingTexture(bool forceBgra = false) + { + _upscaledTexture?.Dispose(); + + var info = new TextureCreateInfo( + _width, + _height, + 1, + 1, + 1, + 1, + 1, + 1, + Format.R8G8B8A8Unorm, + DepthStencilMode.Depth, + Target.Texture2D, + forceBgra ? SwizzleComponent.Blue : SwizzleComponent.Red, + SwizzleComponent.Green, + forceBgra ? SwizzleComponent.Red : SwizzleComponent.Blue, + SwizzleComponent.Alpha); + + _isBgra = forceBgra; + _upscaledTexture = _renderer.CreateTexture(info, 1) as TextureView; + } + + public void SetScalingFilterLevel(float level) + { + _scalingFilterLevel = level; + _updateScalingFilter = true; } } } \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/DescriptorSetUpdater.cs b/Ryujinx.Graphics.Vulkan/DescriptorSetUpdater.cs index 9ac2e61d..19a08502 100644 --- a/Ryujinx.Graphics.Vulkan/DescriptorSetUpdater.cs +++ b/Ryujinx.Graphics.Vulkan/DescriptorSetUpdater.cs @@ -163,6 +163,13 @@ namespace Ryujinx.Graphics.Vulkan SignalDirty(DirtyFlags.Image); } + public void SetImage(int binding, Auto image) + { + _imageRefs[binding] = image; + + SignalDirty(DirtyFlags.Image); + } + public void SetStorageBuffers(CommandBuffer commandBuffer, ReadOnlySpan buffers) { for (int i = 0; i < buffers.Length; i++) diff --git a/Ryujinx.Graphics.Vulkan/Effects/FsrScalingFilter.cs b/Ryujinx.Graphics.Vulkan/Effects/FsrScalingFilter.cs new file mode 100644 index 00000000..a1207059 --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/FsrScalingFilter.cs @@ -0,0 +1,208 @@ +using Ryujinx.Common; +using Ryujinx.Graphics.GAL; +using Ryujinx.Graphics.Shader; +using Ryujinx.Graphics.Shader.Translation; +using Silk.NET.Vulkan; +using System; +using Extent2D = Ryujinx.Graphics.GAL.Extents2D; + +namespace Ryujinx.Graphics.Vulkan.Effects +{ + internal partial class FsrScalingFilter : IScalingFilter + { + private readonly VulkanRenderer _renderer; + private PipelineHelperShader _pipeline; + private ISampler _sampler; + private ShaderCollection _scalingProgram; + private ShaderCollection _sharpeningProgram; + private float _sharpeningLevel = 1; + private Device _device; + private TextureView _intermediaryTexture; + + public float Level + { + get => _sharpeningLevel; + set + { + _sharpeningLevel = MathF.Max(0.01f, value); + } + } + + public FsrScalingFilter(VulkanRenderer renderer, Device device) + { + _device = device; + _renderer = renderer; + + Initialize(); + } + + public void Dispose() + { + _pipeline.Dispose(); + _scalingProgram.Dispose(); + _sharpeningProgram.Dispose(); + _sampler.Dispose(); + _intermediaryTexture?.Dispose(); + } + + public void Initialize() + { + _pipeline = new PipelineHelperShader(_renderer, _device); + + _pipeline.Initialize(); + + var scalingShader = EmbeddedResources.Read("Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrScaling.spv"); + var sharpeningShader = EmbeddedResources.Read("Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrSharpening.spv"); + + var computeBindings = new ShaderBindings( + new[] { 2 }, + Array.Empty(), + new[] { 1 }, + new[] { 0 }); + + var sharpeningBindings = new ShaderBindings( + new[] { 2, 3, 4 }, + Array.Empty(), + new[] { 1 }, + new[] { 0 }); + + _sampler = _renderer.CreateSampler(GAL.SamplerCreateInfo.Create(MinFilter.Linear, MagFilter.Linear)); + + _scalingProgram = _renderer.CreateProgramWithMinimalLayout(new[] + { + new ShaderSource(scalingShader, computeBindings, ShaderStage.Compute, TargetLanguage.Spirv) + }); + + _sharpeningProgram = _renderer.CreateProgramWithMinimalLayout(new[] + { + new ShaderSource(sharpeningShader, sharpeningBindings, ShaderStage.Compute, TargetLanguage.Spirv) + }); + } + + public void Run( + TextureView view, + CommandBufferScoped cbs, + Auto destinationTexture, + Silk.NET.Vulkan.Format format, + int width, + int height, + Extent2D source, + Extent2D destination) + { + if (_intermediaryTexture == null + || _intermediaryTexture.Info.Width != width + || _intermediaryTexture.Info.Height != height + || !_intermediaryTexture.Info.Equals(view.Info)) + { + var originalInfo = view.Info; + + var swapRB = originalInfo.Format.IsBgr() && originalInfo.SwizzleR == SwizzleComponent.Red; + + var info = new TextureCreateInfo( + width, + height, + originalInfo.Depth, + originalInfo.Levels, + originalInfo.Samples, + originalInfo.BlockWidth, + originalInfo.BlockHeight, + originalInfo.BytesPerPixel, + originalInfo.Format, + originalInfo.DepthStencilMode, + originalInfo.Target, + swapRB ? originalInfo.SwizzleB : originalInfo.SwizzleR, + originalInfo.SwizzleG, + swapRB ? originalInfo.SwizzleR : originalInfo.SwizzleB, + originalInfo.SwizzleA); + _intermediaryTexture?.Dispose(); + _intermediaryTexture = _renderer.CreateTexture(info, view.ScaleFactor) as TextureView; + } + + Span viewports = stackalloc GAL.Viewport[1]; + Span> scissors = stackalloc Rectangle[1]; + + viewports[0] = new GAL.Viewport( + new Rectangle(0, 0, view.Width, view.Height), + ViewportSwizzle.PositiveX, + ViewportSwizzle.PositiveY, + ViewportSwizzle.PositiveZ, + ViewportSwizzle.PositiveW, + 0f, + 1f); + + scissors[0] = new Rectangle(0, 0, view.Width, view.Height); + + _pipeline.SetCommandBuffer(cbs); + _pipeline.SetProgram(_scalingProgram); + _pipeline.SetTextureAndSampler(ShaderStage.Compute, 1, view, _sampler); + + float srcWidth = Math.Abs(source.X2 - source.X1); + float srcHeight = Math.Abs(source.Y2 - source.Y1); + float scaleX = srcWidth / view.Width; + float scaleY = srcHeight / view.Height; + + ReadOnlySpan dimensionsBuffer = stackalloc float[] + { + source.X1, + source.X2, + source.Y1, + source.Y2, + destination.X1, + destination.X2, + destination.Y1, + destination.Y2, + scaleX, + scaleY + }; + + int rangeSize = dimensionsBuffer.Length * sizeof(float); + var bufferHandle = _renderer.BufferManager.CreateWithHandle(_renderer, rangeSize, false); + _renderer.BufferManager.SetData(bufferHandle, 0, dimensionsBuffer); + + ReadOnlySpan sharpeningBuffer = stackalloc float[] { 1.5f - (Level * 0.01f * 1.5f)}; + var sharpeningBufferHandle = _renderer.BufferManager.CreateWithHandle(_renderer, sizeof(float), false); + _renderer.BufferManager.SetData(sharpeningBufferHandle, 0, sharpeningBuffer); + + int threadGroupWorkRegionDim = 16; + int dispatchX = (width + (threadGroupWorkRegionDim - 1)) / threadGroupWorkRegionDim; + int dispatchY = (height + (threadGroupWorkRegionDim - 1)) / threadGroupWorkRegionDim; + + var bufferRanges = new BufferRange(bufferHandle, 0, rangeSize); + _pipeline.SetUniformBuffers(stackalloc[] { new BufferAssignment(2, bufferRanges) }); + _pipeline.SetScissors(scissors); + _pipeline.SetViewports(viewports, false); + _pipeline.SetImage(0, _intermediaryTexture, GAL.Format.R8G8B8A8Unorm); + _pipeline.DispatchCompute(dispatchX, dispatchY, 1); + _pipeline.ComputeBarrier(); + + viewports[0] = new GAL.Viewport( + new Rectangle(0, 0, width, height), + ViewportSwizzle.PositiveX, + ViewportSwizzle.PositiveY, + ViewportSwizzle.PositiveZ, + ViewportSwizzle.PositiveW, + 0f, + 1f); + + scissors[0] = new Rectangle(0, 0, width, height); + + // Sharpening pass + _pipeline.SetCommandBuffer(cbs); + _pipeline.SetProgram(_sharpeningProgram); + _pipeline.SetTextureAndSampler(ShaderStage.Compute, 1, _intermediaryTexture, _sampler); + _pipeline.SetUniformBuffers(stackalloc[] { new BufferAssignment(2, bufferRanges) }); + var sharpeningRange = new BufferRange(sharpeningBufferHandle, 0, sizeof(float)); + _pipeline.SetUniformBuffers(stackalloc[] { new BufferAssignment(4, sharpeningRange) }); + _pipeline.SetScissors(scissors); + _pipeline.SetViewports(viewports, false); + _pipeline.SetImage(0, destinationTexture); + _pipeline.DispatchCompute(dispatchX, dispatchY, 1); + _pipeline.ComputeBarrier(); + + _pipeline.Finish(); + + _renderer.BufferManager.Delete(bufferHandle); + _renderer.BufferManager.Delete(sharpeningBufferHandle); + } + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Effects/FxaaPostProcessingEffect.cs b/Ryujinx.Graphics.Vulkan/Effects/FxaaPostProcessingEffect.cs new file mode 100644 index 00000000..0f6a0a7b --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/FxaaPostProcessingEffect.cs @@ -0,0 +1,127 @@ +using Ryujinx.Common; +using Ryujinx.Graphics.GAL; +using Ryujinx.Graphics.Shader; +using Ryujinx.Graphics.Shader.Translation; +using Silk.NET.Vulkan; +using System; + +namespace Ryujinx.Graphics.Vulkan.Effects +{ + internal partial class FxaaPostProcessingEffect : IPostProcessingEffect + { + private readonly VulkanRenderer _renderer; + private ISampler _samplerLinear; + private ShaderCollection _shaderProgram; + + private PipelineHelperShader _pipeline; + private TextureView _texture; + + public FxaaPostProcessingEffect(VulkanRenderer renderer, Device device) + { + _renderer = renderer; + _pipeline = new PipelineHelperShader(renderer, device); + + Initialize(); + } + + public void Dispose() + { + _shaderProgram.Dispose(); + _pipeline.Dispose(); + _samplerLinear.Dispose(); + _texture?.Dispose(); + } + + private void Initialize() + { + _pipeline.Initialize(); + + var shader = EmbeddedResources.Read("Ryujinx.Graphics.Vulkan/Effects/Shaders/Fxaa.spv"); + + var computeBindings = new ShaderBindings( + new[] { 2 }, + Array.Empty(), + new[] { 1 }, + new[] { 0 }); + + _samplerLinear = _renderer.CreateSampler(GAL.SamplerCreateInfo.Create(MinFilter.Linear, MagFilter.Linear)); + + _shaderProgram = _renderer.CreateProgramWithMinimalLayout(new[] + { + new ShaderSource(shader, computeBindings, ShaderStage.Compute, TargetLanguage.Spirv) + }); + } + + public TextureView Run(TextureView view, CommandBufferScoped cbs, int width, int height) + { + if (_texture == null || _texture.Width != view.Width || _texture.Height != view.Height) + { + _texture?.Dispose(); + + var info = view.Info; + + if (view.Info.Format.IsBgr()) + { + info = new TextureCreateInfo(info.Width, + info.Height, + info.Depth, + info.Levels, + info.Samples, + info.BlockWidth, + info.BlockHeight, + info.BytesPerPixel, + info.Format, + info.DepthStencilMode, + info.Target, + info.SwizzleB, + info.SwizzleG, + info.SwizzleR, + info.SwizzleA); + } + _texture = _renderer.CreateTexture(info, view.ScaleFactor) as TextureView; + } + + _pipeline.SetCommandBuffer(cbs); + _pipeline.SetProgram(_shaderProgram); + _pipeline.SetTextureAndSampler(ShaderStage.Compute, 1, view, _samplerLinear); + + ReadOnlySpan resolutionBuffer = stackalloc float[] { view.Width, view.Height }; + int rangeSize = resolutionBuffer.Length * sizeof(float); + var bufferHandle = _renderer.BufferManager.CreateWithHandle(_renderer, rangeSize, false); + + _renderer.BufferManager.SetData(bufferHandle, 0, resolutionBuffer); + + var bufferRanges = new BufferRange(bufferHandle, 0, rangeSize); + _pipeline.SetUniformBuffers(stackalloc[] { new BufferAssignment(2, bufferRanges) }); + + Span viewports = stackalloc GAL.Viewport[1]; + + viewports[0] = new GAL.Viewport( + new Rectangle(0, 0, view.Width, view.Height), + ViewportSwizzle.PositiveX, + ViewportSwizzle.PositiveY, + ViewportSwizzle.PositiveZ, + ViewportSwizzle.PositiveW, + 0f, + 1f); + + Span> scissors = stackalloc Rectangle[1]; + + var dispatchX = BitUtils.DivRoundUp(view.Width, IPostProcessingEffect.LocalGroupSize); + var dispatchY = BitUtils.DivRoundUp(view.Height, IPostProcessingEffect.LocalGroupSize); + + _pipeline.SetScissors(stackalloc[] { new Rectangle(0, 0, view.Width, view.Height) }); + _pipeline.SetViewports(viewports, false); + + _pipeline.SetImage(0, _texture, GAL.Format.R8G8B8A8Unorm); + _pipeline.DispatchCompute(dispatchX, dispatchY, 1); + + _renderer.BufferManager.Delete(bufferHandle); + _pipeline.ComputeBarrier(); + + _pipeline.Finish(); + + return _texture; + } + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Effects/IPostProcessingEffect.cs b/Ryujinx.Graphics.Vulkan/Effects/IPostProcessingEffect.cs new file mode 100644 index 00000000..d36cf01d --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/IPostProcessingEffect.cs @@ -0,0 +1,10 @@ +using System; + +namespace Ryujinx.Graphics.Vulkan.Effects +{ + internal interface IPostProcessingEffect : IDisposable + { + const int LocalGroupSize = 64; + TextureView Run(TextureView view, CommandBufferScoped cbs, int width, int height); + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Effects/IScalingFilter.cs b/Ryujinx.Graphics.Vulkan/Effects/IScalingFilter.cs new file mode 100644 index 00000000..54f809d7 --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/IScalingFilter.cs @@ -0,0 +1,20 @@ +using Silk.NET.Vulkan; +using System; +using Extent2D = Ryujinx.Graphics.GAL.Extents2D; + +namespace Ryujinx.Graphics.Vulkan.Effects +{ + internal interface IScalingFilter : IDisposable + { + float Level { get; set; } + void Run( + TextureView view, + CommandBufferScoped cbs, + Auto destinationTexture, + Format format, + int width, + int height, + Extent2D source, + Extent2D destination); + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrScaling.glsl b/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrScaling.glsl new file mode 100644 index 00000000..5eb74b3d --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrScaling.glsl @@ -0,0 +1,3945 @@ +// Scaling + +#version 430 core +layout (local_size_x = 64) in; +layout( rgba8, binding = 0, set = 3) uniform image2D imgOutput; +layout( binding = 1, set = 2) uniform sampler2D Source; +layout( binding = 2 ) uniform dimensions{ + float srcX0; + float srcX1; + float srcY0; + float srcY1; + float dstX0; + float dstX1; + float dstY0; + float dstY1; + float scaleX; + float scaleY; +}; + +#define A_GPU 1 +#define A_GLSL 1 +//============================================================================================================================== +// +// [A] SHADER PORTABILITY 1.20210629 +// +//============================================================================================================================== +// FidelityFX Super Resolution Sample +// +// Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved. +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files(the "Software"), to deal +// in the Software without restriction, including without limitation the rights +// to use, copy, modify, merge, publish, distribute, sublicense, and / or sell +// copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions : +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN +// THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +// MIT LICENSE +// =========== +// Copyright (c) 2014 Michal Drobot (for concepts used in "FLOAT APPROXIMATIONS"). +// ----------- +// Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation +// files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, +// modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the +// Software is furnished to do so, subject to the following conditions: +// ----------- +// The above copyright notice and this permission notice shall be included in all copies or substantial portions of the +// Software. +// ----------- +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE +// WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR +// COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, +// ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +// ABOUT +// ===== +// Common central point for high-level shading language and C portability for various shader headers. +//------------------------------------------------------------------------------------------------------------------------------ +// DEFINES +// ======= +// A_CPU ..... Include the CPU related code. +// A_GPU ..... Include the GPU related code. +// A_GLSL .... Using GLSL. +// A_HLSL .... Using HLSL. +// A_HLSL_6_2 Using HLSL 6.2 with new 'uint16_t' and related types (requires '-enable-16bit-types'). +// A_NO_16_BIT_CAST Don't use instructions that are not availabe in SPIR-V (needed for running A_HLSL_6_2 on Vulkan) +// A_GCC ..... Using a GCC compatible compiler (else assume MSVC compatible compiler by default). +// ======= +// A_BYTE .... Support 8-bit integer. +// A_HALF .... Support 16-bit integer and floating point. +// A_LONG .... Support 64-bit integer. +// A_DUBL .... Support 64-bit floating point. +// ======= +// A_WAVE .... Support wave-wide operations. +//------------------------------------------------------------------------------------------------------------------------------ +// To get #include "ffx_a.h" working in GLSL use '#extension GL_GOOGLE_include_directive:require'. +//------------------------------------------------------------------------------------------------------------------------------ +// SIMPLIFIED TYPE SYSTEM +// ====================== +// - All ints will be unsigned with exception of when signed is required. +// - Type naming simplified and shortened "A<#components>", +// - H = 16-bit float (half) +// - F = 32-bit float (float) +// - D = 64-bit float (double) +// - P = 1-bit integer (predicate, not using bool because 'B' is used for byte) +// - B = 8-bit integer (byte) +// - W = 16-bit integer (word) +// - U = 32-bit integer (unsigned) +// - L = 64-bit integer (long) +// - Using "AS<#components>" for signed when required. +//------------------------------------------------------------------------------------------------------------------------------ +// TODO +// ==== +// - Make sure 'ALerp*(a,b,m)' does 'b*m+(-a*m+a)' (2 ops). +//------------------------------------------------------------------------------------------------------------------------------ +// CHANGE LOG +// ========== +// 20200914 - Expanded wave ops and prx code. +// 20200713 - Added [ZOL] section, fixed serious bugs in sRGB and Rec.709 color conversion code, etc. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// COMMON +//============================================================================================================================== +#define A_2PI 6.28318530718 +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// CPU +// +// +//============================================================================================================================== +#ifdef A_CPU + // Supporting user defined overrides. + #ifndef A_RESTRICT + #define A_RESTRICT __restrict + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifndef A_STATIC + #define A_STATIC static + #endif +//------------------------------------------------------------------------------------------------------------------------------ + // Same types across CPU and GPU. + // Predicate uses 32-bit integer (C friendly bool). + typedef uint32_t AP1; + typedef float AF1; + typedef double AD1; + typedef uint8_t AB1; + typedef uint16_t AW1; + typedef uint32_t AU1; + typedef uint64_t AL1; + typedef int8_t ASB1; + typedef int16_t ASW1; + typedef int32_t ASU1; + typedef int64_t ASL1; +//------------------------------------------------------------------------------------------------------------------------------ + #define AD1_(a) ((AD1)(a)) + #define AF1_(a) ((AF1)(a)) + #define AL1_(a) ((AL1)(a)) + #define AU1_(a) ((AU1)(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ASL1_(a) ((ASL1)(a)) + #define ASU1_(a) ((ASU1)(a)) +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AU1 AU1_AF1(AF1 a){union{AF1 f;AU1 u;}bits;bits.f=a;return bits.u;} +//------------------------------------------------------------------------------------------------------------------------------ + #define A_TRUE 1 + #define A_FALSE 0 +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// CPU/GPU PORTING +// +//------------------------------------------------------------------------------------------------------------------------------ +// Get CPU and GPU to share all setup code, without duplicate code paths. +// This uses a lower-case prefix for special vector constructs. +// - In C restrict pointers are used. +// - In the shading language, in/inout/out arguments are used. +// This depends on the ability to access a vector value in both languages via array syntax (aka color[2]). +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR ARGUMENT/RETURN/INITIALIZATION PORTABILITY +//============================================================================================================================== + #define retAD2 AD1 *A_RESTRICT + #define retAD3 AD1 *A_RESTRICT + #define retAD4 AD1 *A_RESTRICT + #define retAF2 AF1 *A_RESTRICT + #define retAF3 AF1 *A_RESTRICT + #define retAF4 AF1 *A_RESTRICT + #define retAL2 AL1 *A_RESTRICT + #define retAL3 AL1 *A_RESTRICT + #define retAL4 AL1 *A_RESTRICT + #define retAU2 AU1 *A_RESTRICT + #define retAU3 AU1 *A_RESTRICT + #define retAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define inAD2 AD1 *A_RESTRICT + #define inAD3 AD1 *A_RESTRICT + #define inAD4 AD1 *A_RESTRICT + #define inAF2 AF1 *A_RESTRICT + #define inAF3 AF1 *A_RESTRICT + #define inAF4 AF1 *A_RESTRICT + #define inAL2 AL1 *A_RESTRICT + #define inAL3 AL1 *A_RESTRICT + #define inAL4 AL1 *A_RESTRICT + #define inAU2 AU1 *A_RESTRICT + #define inAU3 AU1 *A_RESTRICT + #define inAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define inoutAD2 AD1 *A_RESTRICT + #define inoutAD3 AD1 *A_RESTRICT + #define inoutAD4 AD1 *A_RESTRICT + #define inoutAF2 AF1 *A_RESTRICT + #define inoutAF3 AF1 *A_RESTRICT + #define inoutAF4 AF1 *A_RESTRICT + #define inoutAL2 AL1 *A_RESTRICT + #define inoutAL3 AL1 *A_RESTRICT + #define inoutAL4 AL1 *A_RESTRICT + #define inoutAU2 AU1 *A_RESTRICT + #define inoutAU3 AU1 *A_RESTRICT + #define inoutAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define outAD2 AD1 *A_RESTRICT + #define outAD3 AD1 *A_RESTRICT + #define outAD4 AD1 *A_RESTRICT + #define outAF2 AF1 *A_RESTRICT + #define outAF3 AF1 *A_RESTRICT + #define outAF4 AF1 *A_RESTRICT + #define outAL2 AL1 *A_RESTRICT + #define outAL3 AL1 *A_RESTRICT + #define outAL4 AL1 *A_RESTRICT + #define outAU2 AU1 *A_RESTRICT + #define outAU3 AU1 *A_RESTRICT + #define outAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define varAD2(x) AD1 x[2] + #define varAD3(x) AD1 x[3] + #define varAD4(x) AD1 x[4] + #define varAF2(x) AF1 x[2] + #define varAF3(x) AF1 x[3] + #define varAF4(x) AF1 x[4] + #define varAL2(x) AL1 x[2] + #define varAL3(x) AL1 x[3] + #define varAL4(x) AL1 x[4] + #define varAU2(x) AU1 x[2] + #define varAU3(x) AU1 x[3] + #define varAU4(x) AU1 x[4] +//------------------------------------------------------------------------------------------------------------------------------ + #define initAD2(x,y) {x,y} + #define initAD3(x,y,z) {x,y,z} + #define initAD4(x,y,z,w) {x,y,z,w} + #define initAF2(x,y) {x,y} + #define initAF3(x,y,z) {x,y,z} + #define initAF4(x,y,z,w) {x,y,z,w} + #define initAL2(x,y) {x,y} + #define initAL3(x,y,z) {x,y,z} + #define initAL4(x,y,z,w) {x,y,z,w} + #define initAU2(x,y) {x,y} + #define initAU3(x,y,z) {x,y,z} + #define initAU4(x,y,z,w) {x,y,z,w} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS +//------------------------------------------------------------------------------------------------------------------------------ +// TODO +// ==== +// - Replace transcendentals with manual versions. +//============================================================================================================================== + #ifdef A_GCC + A_STATIC AD1 AAbsD1(AD1 a){return __builtin_fabs(a);} + A_STATIC AF1 AAbsF1(AF1 a){return __builtin_fabsf(a);} + A_STATIC AU1 AAbsSU1(AU1 a){return AU1_(__builtin_abs(ASU1_(a)));} + A_STATIC AL1 AAbsSL1(AL1 a){return AL1_(__builtin_llabs(ASL1_(a)));} + #else + A_STATIC AD1 AAbsD1(AD1 a){return fabs(a);} + A_STATIC AF1 AAbsF1(AF1 a){return fabsf(a);} + A_STATIC AU1 AAbsSU1(AU1 a){return AU1_(abs(ASU1_(a)));} + A_STATIC AL1 AAbsSL1(AL1 a){return AL1_(labs((long)ASL1_(a)));} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ACosD1(AD1 a){return __builtin_cos(a);} + A_STATIC AF1 ACosF1(AF1 a){return __builtin_cosf(a);} + #else + A_STATIC AD1 ACosD1(AD1 a){return cos(a);} + A_STATIC AF1 ACosF1(AF1 a){return cosf(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ADotD2(inAD2 a,inAD2 b){return a[0]*b[0]+a[1]*b[1];} + A_STATIC AD1 ADotD3(inAD3 a,inAD3 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2];} + A_STATIC AD1 ADotD4(inAD4 a,inAD4 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2]+a[3]*b[3];} + A_STATIC AF1 ADotF2(inAF2 a,inAF2 b){return a[0]*b[0]+a[1]*b[1];} + A_STATIC AF1 ADotF3(inAF3 a,inAF3 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2];} + A_STATIC AF1 ADotF4(inAF4 a,inAF4 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2]+a[3]*b[3];} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 AExp2D1(AD1 a){return __builtin_exp2(a);} + A_STATIC AF1 AExp2F1(AF1 a){return __builtin_exp2f(a);} + #else + A_STATIC AD1 AExp2D1(AD1 a){return exp2(a);} + A_STATIC AF1 AExp2F1(AF1 a){return exp2f(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 AFloorD1(AD1 a){return __builtin_floor(a);} + A_STATIC AF1 AFloorF1(AF1 a){return __builtin_floorf(a);} + #else + A_STATIC AD1 AFloorD1(AD1 a){return floor(a);} + A_STATIC AF1 AFloorF1(AF1 a){return floorf(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ALerpD1(AD1 a,AD1 b,AD1 c){return b*c+(-a*c+a);} + A_STATIC AF1 ALerpF1(AF1 a,AF1 b,AF1 c){return b*c+(-a*c+a);} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ALog2D1(AD1 a){return __builtin_log2(a);} + A_STATIC AF1 ALog2F1(AF1 a){return __builtin_log2f(a);} + #else + A_STATIC AD1 ALog2D1(AD1 a){return log2(a);} + A_STATIC AF1 ALog2F1(AF1 a){return log2f(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 AMaxD1(AD1 a,AD1 b){return a>b?a:b;} + A_STATIC AF1 AMaxF1(AF1 a,AF1 b){return a>b?a:b;} + A_STATIC AL1 AMaxL1(AL1 a,AL1 b){return a>b?a:b;} + A_STATIC AU1 AMaxU1(AU1 a,AU1 b){return a>b?a:b;} +//------------------------------------------------------------------------------------------------------------------------------ + // These follow the convention that A integer types don't have signage, until they are operated on. + A_STATIC AL1 AMaxSL1(AL1 a,AL1 b){return (ASL1_(a)>ASL1_(b))?a:b;} + A_STATIC AU1 AMaxSU1(AU1 a,AU1 b){return (ASU1_(a)>ASU1_(b))?a:b;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 AMinD1(AD1 a,AD1 b){return a>ASL1_(b));} + A_STATIC AU1 AShrSU1(AU1 a,AU1 b){return AU1_(ASU1_(a)>>ASU1_(b));} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ASinD1(AD1 a){return __builtin_sin(a);} + A_STATIC AF1 ASinF1(AF1 a){return __builtin_sinf(a);} + #else + A_STATIC AD1 ASinD1(AD1 a){return sin(a);} + A_STATIC AF1 ASinF1(AF1 a){return sinf(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ASqrtD1(AD1 a){return __builtin_sqrt(a);} + A_STATIC AF1 ASqrtF1(AF1 a){return __builtin_sqrtf(a);} + #else + A_STATIC AD1 ASqrtD1(AD1 a){return sqrt(a);} + A_STATIC AF1 ASqrtF1(AF1 a){return sqrtf(a);} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS - DEPENDENT +//============================================================================================================================== + A_STATIC AD1 AClampD1(AD1 x,AD1 n,AD1 m){return AMaxD1(n,AMinD1(x,m));} + A_STATIC AF1 AClampF1(AF1 x,AF1 n,AF1 m){return AMaxF1(n,AMinF1(x,m));} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 AFractD1(AD1 a){return a-AFloorD1(a);} + A_STATIC AF1 AFractF1(AF1 a){return a-AFloorF1(a);} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 APowD1(AD1 a,AD1 b){return AExp2D1(b*ALog2D1(a));} + A_STATIC AF1 APowF1(AF1 a,AF1 b){return AExp2F1(b*ALog2F1(a));} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ARsqD1(AD1 a){return ARcpD1(ASqrtD1(a));} + A_STATIC AF1 ARsqF1(AF1 a){return ARcpF1(ASqrtF1(a));} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ASatD1(AD1 a){return AMinD1(1.0,AMaxD1(0.0,a));} + A_STATIC AF1 ASatF1(AF1 a){return AMinF1(1.0f,AMaxF1(0.0f,a));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR OPS +//------------------------------------------------------------------------------------------------------------------------------ +// These are added as needed for production or prototyping, so not necessarily a complete set. +// They follow a convention of taking in a destination and also returning the destination value to increase utility. +//============================================================================================================================== + A_STATIC retAD2 opAAbsD2(outAD2 d,inAD2 a){d[0]=AAbsD1(a[0]);d[1]=AAbsD1(a[1]);return d;} + A_STATIC retAD3 opAAbsD3(outAD3 d,inAD3 a){d[0]=AAbsD1(a[0]);d[1]=AAbsD1(a[1]);d[2]=AAbsD1(a[2]);return d;} + A_STATIC retAD4 opAAbsD4(outAD4 d,inAD4 a){d[0]=AAbsD1(a[0]);d[1]=AAbsD1(a[1]);d[2]=AAbsD1(a[2]);d[3]=AAbsD1(a[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAAbsF2(outAF2 d,inAF2 a){d[0]=AAbsF1(a[0]);d[1]=AAbsF1(a[1]);return d;} + A_STATIC retAF3 opAAbsF3(outAF3 d,inAF3 a){d[0]=AAbsF1(a[0]);d[1]=AAbsF1(a[1]);d[2]=AAbsF1(a[2]);return d;} + A_STATIC retAF4 opAAbsF4(outAF4 d,inAF4 a){d[0]=AAbsF1(a[0]);d[1]=AAbsF1(a[1]);d[2]=AAbsF1(a[2]);d[3]=AAbsF1(a[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAAddD2(outAD2 d,inAD2 a,inAD2 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];return d;} + A_STATIC retAD3 opAAddD3(outAD3 d,inAD3 a,inAD3 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];return d;} + A_STATIC retAD4 opAAddD4(outAD4 d,inAD4 a,inAD4 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];d[3]=a[3]+b[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAAddF2(outAF2 d,inAF2 a,inAF2 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];return d;} + A_STATIC retAF3 opAAddF3(outAF3 d,inAF3 a,inAF3 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];return d;} + A_STATIC retAF4 opAAddF4(outAF4 d,inAF4 a,inAF4 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];d[3]=a[3]+b[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opAAddOneD2(outAD2 d,inAD2 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;return d;} + A_STATIC retAD3 opAAddOneD3(outAD3 d,inAD3 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;return d;} + A_STATIC retAD4 opAAddOneD4(outAD4 d,inAD4 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;d[3]=a[3]+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAAddOneF2(outAF2 d,inAF2 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;return d;} + A_STATIC retAF3 opAAddOneF3(outAF3 d,inAF3 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;return d;} + A_STATIC retAF4 opAAddOneF4(outAF4 d,inAF4 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;d[3]=a[3]+b;return d;} +//============================================================================================================================== + A_STATIC retAD2 opACpyD2(outAD2 d,inAD2 a){d[0]=a[0];d[1]=a[1];return d;} + A_STATIC retAD3 opACpyD3(outAD3 d,inAD3 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];return d;} + A_STATIC retAD4 opACpyD4(outAD4 d,inAD4 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];d[3]=a[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opACpyF2(outAF2 d,inAF2 a){d[0]=a[0];d[1]=a[1];return d;} + A_STATIC retAF3 opACpyF3(outAF3 d,inAF3 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];return d;} + A_STATIC retAF4 opACpyF4(outAF4 d,inAF4 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];d[3]=a[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opALerpD2(outAD2 d,inAD2 a,inAD2 b,inAD2 c){d[0]=ALerpD1(a[0],b[0],c[0]);d[1]=ALerpD1(a[1],b[1],c[1]);return d;} + A_STATIC retAD3 opALerpD3(outAD3 d,inAD3 a,inAD3 b,inAD3 c){d[0]=ALerpD1(a[0],b[0],c[0]);d[1]=ALerpD1(a[1],b[1],c[1]);d[2]=ALerpD1(a[2],b[2],c[2]);return d;} + A_STATIC retAD4 opALerpD4(outAD4 d,inAD4 a,inAD4 b,inAD4 c){d[0]=ALerpD1(a[0],b[0],c[0]);d[1]=ALerpD1(a[1],b[1],c[1]);d[2]=ALerpD1(a[2],b[2],c[2]);d[3]=ALerpD1(a[3],b[3],c[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opALerpF2(outAF2 d,inAF2 a,inAF2 b,inAF2 c){d[0]=ALerpF1(a[0],b[0],c[0]);d[1]=ALerpF1(a[1],b[1],c[1]);return d;} + A_STATIC retAF3 opALerpF3(outAF3 d,inAF3 a,inAF3 b,inAF3 c){d[0]=ALerpF1(a[0],b[0],c[0]);d[1]=ALerpF1(a[1],b[1],c[1]);d[2]=ALerpF1(a[2],b[2],c[2]);return d;} + A_STATIC retAF4 opALerpF4(outAF4 d,inAF4 a,inAF4 b,inAF4 c){d[0]=ALerpF1(a[0],b[0],c[0]);d[1]=ALerpF1(a[1],b[1],c[1]);d[2]=ALerpF1(a[2],b[2],c[2]);d[3]=ALerpF1(a[3],b[3],c[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opALerpOneD2(outAD2 d,inAD2 a,inAD2 b,AD1 c){d[0]=ALerpD1(a[0],b[0],c);d[1]=ALerpD1(a[1],b[1],c);return d;} + A_STATIC retAD3 opALerpOneD3(outAD3 d,inAD3 a,inAD3 b,AD1 c){d[0]=ALerpD1(a[0],b[0],c);d[1]=ALerpD1(a[1],b[1],c);d[2]=ALerpD1(a[2],b[2],c);return d;} + A_STATIC retAD4 opALerpOneD4(outAD4 d,inAD4 a,inAD4 b,AD1 c){d[0]=ALerpD1(a[0],b[0],c);d[1]=ALerpD1(a[1],b[1],c);d[2]=ALerpD1(a[2],b[2],c);d[3]=ALerpD1(a[3],b[3],c);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opALerpOneF2(outAF2 d,inAF2 a,inAF2 b,AF1 c){d[0]=ALerpF1(a[0],b[0],c);d[1]=ALerpF1(a[1],b[1],c);return d;} + A_STATIC retAF3 opALerpOneF3(outAF3 d,inAF3 a,inAF3 b,AF1 c){d[0]=ALerpF1(a[0],b[0],c);d[1]=ALerpF1(a[1],b[1],c);d[2]=ALerpF1(a[2],b[2],c);return d;} + A_STATIC retAF4 opALerpOneF4(outAF4 d,inAF4 a,inAF4 b,AF1 c){d[0]=ALerpF1(a[0],b[0],c);d[1]=ALerpF1(a[1],b[1],c);d[2]=ALerpF1(a[2],b[2],c);d[3]=ALerpF1(a[3],b[3],c);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMaxD2(outAD2 d,inAD2 a,inAD2 b){d[0]=AMaxD1(a[0],b[0]);d[1]=AMaxD1(a[1],b[1]);return d;} + A_STATIC retAD3 opAMaxD3(outAD3 d,inAD3 a,inAD3 b){d[0]=AMaxD1(a[0],b[0]);d[1]=AMaxD1(a[1],b[1]);d[2]=AMaxD1(a[2],b[2]);return d;} + A_STATIC retAD4 opAMaxD4(outAD4 d,inAD4 a,inAD4 b){d[0]=AMaxD1(a[0],b[0]);d[1]=AMaxD1(a[1],b[1]);d[2]=AMaxD1(a[2],b[2]);d[3]=AMaxD1(a[3],b[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMaxF2(outAF2 d,inAF2 a,inAF2 b){d[0]=AMaxF1(a[0],b[0]);d[1]=AMaxF1(a[1],b[1]);return d;} + A_STATIC retAF3 opAMaxF3(outAF3 d,inAF3 a,inAF3 b){d[0]=AMaxF1(a[0],b[0]);d[1]=AMaxF1(a[1],b[1]);d[2]=AMaxF1(a[2],b[2]);return d;} + A_STATIC retAF4 opAMaxF4(outAF4 d,inAF4 a,inAF4 b){d[0]=AMaxF1(a[0],b[0]);d[1]=AMaxF1(a[1],b[1]);d[2]=AMaxF1(a[2],b[2]);d[3]=AMaxF1(a[3],b[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMinD2(outAD2 d,inAD2 a,inAD2 b){d[0]=AMinD1(a[0],b[0]);d[1]=AMinD1(a[1],b[1]);return d;} + A_STATIC retAD3 opAMinD3(outAD3 d,inAD3 a,inAD3 b){d[0]=AMinD1(a[0],b[0]);d[1]=AMinD1(a[1],b[1]);d[2]=AMinD1(a[2],b[2]);return d;} + A_STATIC retAD4 opAMinD4(outAD4 d,inAD4 a,inAD4 b){d[0]=AMinD1(a[0],b[0]);d[1]=AMinD1(a[1],b[1]);d[2]=AMinD1(a[2],b[2]);d[3]=AMinD1(a[3],b[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMinF2(outAF2 d,inAF2 a,inAF2 b){d[0]=AMinF1(a[0],b[0]);d[1]=AMinF1(a[1],b[1]);return d;} + A_STATIC retAF3 opAMinF3(outAF3 d,inAF3 a,inAF3 b){d[0]=AMinF1(a[0],b[0]);d[1]=AMinF1(a[1],b[1]);d[2]=AMinF1(a[2],b[2]);return d;} + A_STATIC retAF4 opAMinF4(outAF4 d,inAF4 a,inAF4 b){d[0]=AMinF1(a[0],b[0]);d[1]=AMinF1(a[1],b[1]);d[2]=AMinF1(a[2],b[2]);d[3]=AMinF1(a[3],b[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMulD2(outAD2 d,inAD2 a,inAD2 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];return d;} + A_STATIC retAD3 opAMulD3(outAD3 d,inAD3 a,inAD3 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];return d;} + A_STATIC retAD4 opAMulD4(outAD4 d,inAD4 a,inAD4 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];d[3]=a[3]*b[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMulF2(outAF2 d,inAF2 a,inAF2 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];return d;} + A_STATIC retAF3 opAMulF3(outAF3 d,inAF3 a,inAF3 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];return d;} + A_STATIC retAF4 opAMulF4(outAF4 d,inAF4 a,inAF4 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];d[3]=a[3]*b[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMulOneD2(outAD2 d,inAD2 a,AD1 b){d[0]=a[0]*b;d[1]=a[1]*b;return d;} + A_STATIC retAD3 opAMulOneD3(outAD3 d,inAD3 a,AD1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;return d;} + A_STATIC retAD4 opAMulOneD4(outAD4 d,inAD4 a,AD1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;d[3]=a[3]*b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMulOneF2(outAF2 d,inAF2 a,AF1 b){d[0]=a[0]*b;d[1]=a[1]*b;return d;} + A_STATIC retAF3 opAMulOneF3(outAF3 d,inAF3 a,AF1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;return d;} + A_STATIC retAF4 opAMulOneF4(outAF4 d,inAF4 a,AF1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;d[3]=a[3]*b;return d;} +//============================================================================================================================== + A_STATIC retAD2 opANegD2(outAD2 d,inAD2 a){d[0]=-a[0];d[1]=-a[1];return d;} + A_STATIC retAD3 opANegD3(outAD3 d,inAD3 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];return d;} + A_STATIC retAD4 opANegD4(outAD4 d,inAD4 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];d[3]=-a[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opANegF2(outAF2 d,inAF2 a){d[0]=-a[0];d[1]=-a[1];return d;} + A_STATIC retAF3 opANegF3(outAF3 d,inAF3 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];return d;} + A_STATIC retAF4 opANegF4(outAF4 d,inAF4 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];d[3]=-a[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opARcpD2(outAD2 d,inAD2 a){d[0]=ARcpD1(a[0]);d[1]=ARcpD1(a[1]);return d;} + A_STATIC retAD3 opARcpD3(outAD3 d,inAD3 a){d[0]=ARcpD1(a[0]);d[1]=ARcpD1(a[1]);d[2]=ARcpD1(a[2]);return d;} + A_STATIC retAD4 opARcpD4(outAD4 d,inAD4 a){d[0]=ARcpD1(a[0]);d[1]=ARcpD1(a[1]);d[2]=ARcpD1(a[2]);d[3]=ARcpD1(a[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opARcpF2(outAF2 d,inAF2 a){d[0]=ARcpF1(a[0]);d[1]=ARcpF1(a[1]);return d;} + A_STATIC retAF3 opARcpF3(outAF3 d,inAF3 a){d[0]=ARcpF1(a[0]);d[1]=ARcpF1(a[1]);d[2]=ARcpF1(a[2]);return d;} + A_STATIC retAF4 opARcpF4(outAF4 d,inAF4 a){d[0]=ARcpF1(a[0]);d[1]=ARcpF1(a[1]);d[2]=ARcpF1(a[2]);d[3]=ARcpF1(a[3]);return d;} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HALF FLOAT PACKING +//============================================================================================================================== + // Convert float to half (in lower 16-bits of output). + // Same fast technique as documented here: ftp://ftp.fox-toolkit.org/pub/fasthalffloatconversion.pdf + // Supports denormals. + // Conversion rules are to make computations possibly "safer" on the GPU, + // -INF & -NaN -> -65504 + // +INF & +NaN -> +65504 + A_STATIC AU1 AU1_AH1_AF1(AF1 f){ + static AW1 base[512]={ + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0001,0x0002,0x0004,0x0008,0x0010,0x0020,0x0040,0x0080,0x0100, + 0x0200,0x0400,0x0800,0x0c00,0x1000,0x1400,0x1800,0x1c00,0x2000,0x2400,0x2800,0x2c00,0x3000,0x3400,0x3800,0x3c00, + 0x4000,0x4400,0x4800,0x4c00,0x5000,0x5400,0x5800,0x5c00,0x6000,0x6400,0x6800,0x6c00,0x7000,0x7400,0x7800,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8001,0x8002,0x8004,0x8008,0x8010,0x8020,0x8040,0x8080,0x8100, + 0x8200,0x8400,0x8800,0x8c00,0x9000,0x9400,0x9800,0x9c00,0xa000,0xa400,0xa800,0xac00,0xb000,0xb400,0xb800,0xbc00, + 0xc000,0xc400,0xc800,0xcc00,0xd000,0xd400,0xd800,0xdc00,0xe000,0xe400,0xe800,0xec00,0xf000,0xf400,0xf800,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff}; + static AB1 shift[512]={ + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x17,0x16,0x15,0x14,0x13,0x12,0x11,0x10,0x0f, + 0x0e,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d, + 0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x17,0x16,0x15,0x14,0x13,0x12,0x11,0x10,0x0f, + 0x0e,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d, + 0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18}; + union{AF1 f;AU1 u;}bits;bits.f=f;AU1 u=bits.u;AU1 i=u>>23;return (AU1)(base[i])+((u&0x7fffff)>>shift[i]);} +//------------------------------------------------------------------------------------------------------------------------------ + // Used to output packed constant. + A_STATIC AU1 AU1_AH2_AF2(inAF2 a){return AU1_AH1_AF1(a[0])+(AU1_AH1_AF1(a[1])<<16);} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// GLSL +// +// +//============================================================================================================================== +#if defined(A_GLSL) && defined(A_GPU) + #ifndef A_SKIP_EXT + #ifdef A_HALF + #extension GL_EXT_shader_16bit_storage:require + #extension GL_EXT_shader_explicit_arithmetic_types:require + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_LONG + #extension GL_ARB_gpu_shader_int64:require + #extension GL_NV_shader_atomic_int64:require + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_WAVE + #extension GL_KHR_shader_subgroup_arithmetic:require + #extension GL_KHR_shader_subgroup_ballot:require + #extension GL_KHR_shader_subgroup_quad:require + #extension GL_KHR_shader_subgroup_shuffle:require + #endif + #endif +//============================================================================================================================== + #define AP1 bool + #define AP2 bvec2 + #define AP3 bvec3 + #define AP4 bvec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF1 float + #define AF2 vec2 + #define AF3 vec3 + #define AF4 vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1 uint + #define AU2 uvec2 + #define AU3 uvec3 + #define AU4 uvec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASU1 int + #define ASU2 ivec2 + #define ASU3 ivec3 + #define ASU4 ivec4 +//============================================================================================================================== + #define AF1_AU1(x) uintBitsToFloat(AU1(x)) + #define AF2_AU2(x) uintBitsToFloat(AU2(x)) + #define AF3_AU3(x) uintBitsToFloat(AU3(x)) + #define AF4_AU4(x) uintBitsToFloat(AU4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AF1(x) floatBitsToUint(AF1(x)) + #define AU2_AF2(x) floatBitsToUint(AF2(x)) + #define AU3_AF3(x) floatBitsToUint(AF3(x)) + #define AU4_AF4(x) floatBitsToUint(AF4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH1_AF1_x(AF1 a){return packHalf2x16(AF2(a,0.0));} + #define AU1_AH1_AF1(a) AU1_AH1_AF1_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AH2_AF2 packHalf2x16 + #define AU1_AW2Unorm_AF2 packUnorm2x16 + #define AU1_AB4Unorm_AF4 packUnorm4x8 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF2_AH2_AU1 unpackHalf2x16 + #define AF2_AW2Unorm_AU1 unpackUnorm2x16 + #define AF4_AB4Unorm_AU1 unpackUnorm4x8 +//============================================================================================================================== + AF1 AF1_x(AF1 a){return AF1(a);} + AF2 AF2_x(AF1 a){return AF2(a,a);} + AF3 AF3_x(AF1 a){return AF3(a,a,a);} + AF4 AF4_x(AF1 a){return AF4(a,a,a,a);} + #define AF1_(a) AF1_x(AF1(a)) + #define AF2_(a) AF2_x(AF1(a)) + #define AF3_(a) AF3_x(AF1(a)) + #define AF4_(a) AF4_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_x(AU1 a){return AU1(a);} + AU2 AU2_x(AU1 a){return AU2(a,a);} + AU3 AU3_x(AU1 a){return AU3(a,a,a);} + AU4 AU4_x(AU1 a){return AU4(a,a,a,a);} + #define AU1_(a) AU1_x(AU1(a)) + #define AU2_(a) AU2_x(AU1(a)) + #define AU3_(a) AU3_x(AU1(a)) + #define AU4_(a) AU4_x(AU1(a)) +//============================================================================================================================== + AU1 AAbsSU1(AU1 a){return AU1(abs(ASU1(a)));} + AU2 AAbsSU2(AU2 a){return AU2(abs(ASU2(a)));} + AU3 AAbsSU3(AU3 a){return AU3(abs(ASU3(a)));} + AU4 AAbsSU4(AU4 a){return AU4(abs(ASU4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 ABfe(AU1 src,AU1 off,AU1 bits){return bitfieldExtract(src,ASU1(off),ASU1(bits));} + AU1 ABfi(AU1 src,AU1 ins,AU1 mask){return (ins&mask)|(src&(~mask));} + // Proxy for V_BFI_B32 where the 'mask' is set as 'bits', 'mask=(1<>ASU1(b));} + AU2 AShrSU2(AU2 a,AU2 b){return AU2(ASU2(a)>>ASU2(b));} + AU3 AShrSU3(AU3 a,AU3 b){return AU3(ASU3(a)>>ASU3(b));} + AU4 AShrSU4(AU4 a,AU4 b){return AU4(ASU4(a)>>ASU4(b));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL BYTE +//============================================================================================================================== + #ifdef A_BYTE + #define AB1 uint8_t + #define AB2 u8vec2 + #define AB3 u8vec3 + #define AB4 u8vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASB1 int8_t + #define ASB2 i8vec2 + #define ASB3 i8vec3 + #define ASB4 i8vec4 +//------------------------------------------------------------------------------------------------------------------------------ + AB1 AB1_x(AB1 a){return AB1(a);} + AB2 AB2_x(AB1 a){return AB2(a,a);} + AB3 AB3_x(AB1 a){return AB3(a,a,a);} + AB4 AB4_x(AB1 a){return AB4(a,a,a,a);} + #define AB1_(a) AB1_x(AB1(a)) + #define AB2_(a) AB2_x(AB1(a)) + #define AB3_(a) AB3_x(AB1(a)) + #define AB4_(a) AB4_x(AB1(a)) + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL HALF +//============================================================================================================================== + #ifdef A_HALF + #define AH1 float16_t + #define AH2 f16vec2 + #define AH3 f16vec3 + #define AH4 f16vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AW1 uint16_t + #define AW2 u16vec2 + #define AW3 u16vec3 + #define AW4 u16vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASW1 int16_t + #define ASW2 i16vec2 + #define ASW3 i16vec3 + #define ASW4 i16vec4 +//============================================================================================================================== + #define AH2_AU1(x) unpackFloat2x16(AU1(x)) + AH4 AH4_AU2_x(AU2 x){return AH4(unpackFloat2x16(x.x),unpackFloat2x16(x.y));} + #define AH4_AU2(x) AH4_AU2_x(AU2(x)) + #define AW2_AU1(x) unpackUint2x16(AU1(x)) + #define AW4_AU2(x) unpackUint4x16(pack64(AU2(x))) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AH2(x) packFloat2x16(AH2(x)) + AU2 AU2_AH4_x(AH4 x){return AU2(packFloat2x16(x.xy),packFloat2x16(x.zw));} + #define AU2_AH4(x) AU2_AH4_x(AH4(x)) + #define AU1_AW2(x) packUint2x16(AW2(x)) + #define AU2_AW4(x) unpack32(packUint4x16(AW4(x))) +//============================================================================================================================== + #define AW1_AH1(x) halfBitsToUint16(AH1(x)) + #define AW2_AH2(x) halfBitsToUint16(AH2(x)) + #define AW3_AH3(x) halfBitsToUint16(AH3(x)) + #define AW4_AH4(x) halfBitsToUint16(AH4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AH1_AW1(x) uint16BitsToHalf(AW1(x)) + #define AH2_AW2(x) uint16BitsToHalf(AW2(x)) + #define AH3_AW3(x) uint16BitsToHalf(AW3(x)) + #define AH4_AW4(x) uint16BitsToHalf(AW4(x)) +//============================================================================================================================== + AH1 AH1_x(AH1 a){return AH1(a);} + AH2 AH2_x(AH1 a){return AH2(a,a);} + AH3 AH3_x(AH1 a){return AH3(a,a,a);} + AH4 AH4_x(AH1 a){return AH4(a,a,a,a);} + #define AH1_(a) AH1_x(AH1(a)) + #define AH2_(a) AH2_x(AH1(a)) + #define AH3_(a) AH3_x(AH1(a)) + #define AH4_(a) AH4_x(AH1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AW1_x(AW1 a){return AW1(a);} + AW2 AW2_x(AW1 a){return AW2(a,a);} + AW3 AW3_x(AW1 a){return AW3(a,a,a);} + AW4 AW4_x(AW1 a){return AW4(a,a,a,a);} + #define AW1_(a) AW1_x(AW1(a)) + #define AW2_(a) AW2_x(AW1(a)) + #define AW3_(a) AW3_x(AW1(a)) + #define AW4_(a) AW4_x(AW1(a)) +//============================================================================================================================== + AW1 AAbsSW1(AW1 a){return AW1(abs(ASW1(a)));} + AW2 AAbsSW2(AW2 a){return AW2(abs(ASW2(a)));} + AW3 AAbsSW3(AW3 a){return AW3(abs(ASW3(a)));} + AW4 AAbsSW4(AW4 a){return AW4(abs(ASW4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AClampH1(AH1 x,AH1 n,AH1 m){return clamp(x,n,m);} + AH2 AClampH2(AH2 x,AH2 n,AH2 m){return clamp(x,n,m);} + AH3 AClampH3(AH3 x,AH3 n,AH3 m){return clamp(x,n,m);} + AH4 AClampH4(AH4 x,AH4 n,AH4 m){return clamp(x,n,m);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFractH1(AH1 x){return fract(x);} + AH2 AFractH2(AH2 x){return fract(x);} + AH3 AFractH3(AH3 x){return fract(x);} + AH4 AFractH4(AH4 x){return fract(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ALerpH1(AH1 x,AH1 y,AH1 a){return mix(x,y,a);} + AH2 ALerpH2(AH2 x,AH2 y,AH2 a){return mix(x,y,a);} + AH3 ALerpH3(AH3 x,AH3 y,AH3 a){return mix(x,y,a);} + AH4 ALerpH4(AH4 x,AH4 y,AH4 a){return mix(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + // No packed version of max3. + AH1 AMax3H1(AH1 x,AH1 y,AH1 z){return max(x,max(y,z));} + AH2 AMax3H2(AH2 x,AH2 y,AH2 z){return max(x,max(y,z));} + AH3 AMax3H3(AH3 x,AH3 y,AH3 z){return max(x,max(y,z));} + AH4 AMax3H4(AH4 x,AH4 y,AH4 z){return max(x,max(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMaxSW1(AW1 a,AW1 b){return AW1(max(ASU1(a),ASU1(b)));} + AW2 AMaxSW2(AW2 a,AW2 b){return AW2(max(ASU2(a),ASU2(b)));} + AW3 AMaxSW3(AW3 a,AW3 b){return AW3(max(ASU3(a),ASU3(b)));} + AW4 AMaxSW4(AW4 a,AW4 b){return AW4(max(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + // No packed version of min3. + AH1 AMin3H1(AH1 x,AH1 y,AH1 z){return min(x,min(y,z));} + AH2 AMin3H2(AH2 x,AH2 y,AH2 z){return min(x,min(y,z));} + AH3 AMin3H3(AH3 x,AH3 y,AH3 z){return min(x,min(y,z));} + AH4 AMin3H4(AH4 x,AH4 y,AH4 z){return min(x,min(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMinSW1(AW1 a,AW1 b){return AW1(min(ASU1(a),ASU1(b)));} + AW2 AMinSW2(AW2 a,AW2 b){return AW2(min(ASU2(a),ASU2(b)));} + AW3 AMinSW3(AW3 a,AW3 b){return AW3(min(ASU3(a),ASU3(b)));} + AW4 AMinSW4(AW4 a,AW4 b){return AW4(min(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARcpH1(AH1 x){return AH1_(1.0)/x;} + AH2 ARcpH2(AH2 x){return AH2_(1.0)/x;} + AH3 ARcpH3(AH3 x){return AH3_(1.0)/x;} + AH4 ARcpH4(AH4 x){return AH4_(1.0)/x;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARsqH1(AH1 x){return AH1_(1.0)/sqrt(x);} + AH2 ARsqH2(AH2 x){return AH2_(1.0)/sqrt(x);} + AH3 ARsqH3(AH3 x){return AH3_(1.0)/sqrt(x);} + AH4 ARsqH4(AH4 x){return AH4_(1.0)/sqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASatH1(AH1 x){return clamp(x,AH1_(0.0),AH1_(1.0));} + AH2 ASatH2(AH2 x){return clamp(x,AH2_(0.0),AH2_(1.0));} + AH3 ASatH3(AH3 x){return clamp(x,AH3_(0.0),AH3_(1.0));} + AH4 ASatH4(AH4 x){return clamp(x,AH4_(0.0),AH4_(1.0));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AShrSW1(AW1 a,AW1 b){return AW1(ASW1(a)>>ASW1(b));} + AW2 AShrSW2(AW2 a,AW2 b){return AW2(ASW2(a)>>ASW2(b));} + AW3 AShrSW3(AW3 a,AW3 b){return AW3(ASW3(a)>>ASW3(b));} + AW4 AShrSW4(AW4 a,AW4 b){return AW4(ASW4(a)>>ASW4(b));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL DOUBLE +//============================================================================================================================== + #ifdef A_DUBL + #define AD1 double + #define AD2 dvec2 + #define AD3 dvec3 + #define AD4 dvec4 +//------------------------------------------------------------------------------------------------------------------------------ + AD1 AD1_x(AD1 a){return AD1(a);} + AD2 AD2_x(AD1 a){return AD2(a,a);} + AD3 AD3_x(AD1 a){return AD3(a,a,a);} + AD4 AD4_x(AD1 a){return AD4(a,a,a,a);} + #define AD1_(a) AD1_x(AD1(a)) + #define AD2_(a) AD2_x(AD1(a)) + #define AD3_(a) AD3_x(AD1(a)) + #define AD4_(a) AD4_x(AD1(a)) +//============================================================================================================================== + AD1 AFractD1(AD1 x){return fract(x);} + AD2 AFractD2(AD2 x){return fract(x);} + AD3 AFractD3(AD3 x){return fract(x);} + AD4 AFractD4(AD4 x){return fract(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ALerpD1(AD1 x,AD1 y,AD1 a){return mix(x,y,a);} + AD2 ALerpD2(AD2 x,AD2 y,AD2 a){return mix(x,y,a);} + AD3 ALerpD3(AD3 x,AD3 y,AD3 a){return mix(x,y,a);} + AD4 ALerpD4(AD4 x,AD4 y,AD4 a){return mix(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARcpD1(AD1 x){return AD1_(1.0)/x;} + AD2 ARcpD2(AD2 x){return AD2_(1.0)/x;} + AD3 ARcpD3(AD3 x){return AD3_(1.0)/x;} + AD4 ARcpD4(AD4 x){return AD4_(1.0)/x;} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARsqD1(AD1 x){return AD1_(1.0)/sqrt(x);} + AD2 ARsqD2(AD2 x){return AD2_(1.0)/sqrt(x);} + AD3 ARsqD3(AD3 x){return AD3_(1.0)/sqrt(x);} + AD4 ARsqD4(AD4 x){return AD4_(1.0)/sqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ASatD1(AD1 x){return clamp(x,AD1_(0.0),AD1_(1.0));} + AD2 ASatD2(AD2 x){return clamp(x,AD2_(0.0),AD2_(1.0));} + AD3 ASatD3(AD3 x){return clamp(x,AD3_(0.0),AD3_(1.0));} + AD4 ASatD4(AD4 x){return clamp(x,AD4_(0.0),AD4_(1.0));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL LONG +//============================================================================================================================== + #ifdef A_LONG + #define AL1 uint64_t + #define AL2 u64vec2 + #define AL3 u64vec3 + #define AL4 u64vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASL1 int64_t + #define ASL2 i64vec2 + #define ASL3 i64vec3 + #define ASL4 i64vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AL1_AU2(x) packUint2x32(AU2(x)) + #define AU2_AL1(x) unpackUint2x32(AL1(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AL1 AL1_x(AL1 a){return AL1(a);} + AL2 AL2_x(AL1 a){return AL2(a,a);} + AL3 AL3_x(AL1 a){return AL3(a,a,a);} + AL4 AL4_x(AL1 a){return AL4(a,a,a,a);} + #define AL1_(a) AL1_x(AL1(a)) + #define AL2_(a) AL2_x(AL1(a)) + #define AL3_(a) AL3_x(AL1(a)) + #define AL4_(a) AL4_x(AL1(a)) +//============================================================================================================================== + AL1 AAbsSL1(AL1 a){return AL1(abs(ASL1(a)));} + AL2 AAbsSL2(AL2 a){return AL2(abs(ASL2(a)));} + AL3 AAbsSL3(AL3 a){return AL3(abs(ASL3(a)));} + AL4 AAbsSL4(AL4 a){return AL4(abs(ASL4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AL1 AMaxSL1(AL1 a,AL1 b){return AL1(max(ASU1(a),ASU1(b)));} + AL2 AMaxSL2(AL2 a,AL2 b){return AL2(max(ASU2(a),ASU2(b)));} + AL3 AMaxSL3(AL3 a,AL3 b){return AL3(max(ASU3(a),ASU3(b)));} + AL4 AMaxSL4(AL4 a,AL4 b){return AL4(max(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AL1 AMinSL1(AL1 a,AL1 b){return AL1(min(ASU1(a),ASU1(b)));} + AL2 AMinSL2(AL2 a,AL2 b){return AL2(min(ASU2(a),ASU2(b)));} + AL3 AMinSL3(AL3 a,AL3 b){return AL3(min(ASU3(a),ASU3(b)));} + AL4 AMinSL4(AL4 a,AL4 b){return AL4(min(ASU4(a),ASU4(b)));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// WAVE OPERATIONS +//============================================================================================================================== + #ifdef A_WAVE + // Where 'x' must be a compile time literal. + AF1 AWaveXorF1(AF1 v,AU1 x){return subgroupShuffleXor(v,x);} + AF2 AWaveXorF2(AF2 v,AU1 x){return subgroupShuffleXor(v,x);} + AF3 AWaveXorF3(AF3 v,AU1 x){return subgroupShuffleXor(v,x);} + AF4 AWaveXorF4(AF4 v,AU1 x){return subgroupShuffleXor(v,x);} + AU1 AWaveXorU1(AU1 v,AU1 x){return subgroupShuffleXor(v,x);} + AU2 AWaveXorU2(AU2 v,AU1 x){return subgroupShuffleXor(v,x);} + AU3 AWaveXorU3(AU3 v,AU1 x){return subgroupShuffleXor(v,x);} + AU4 AWaveXorU4(AU4 v,AU1 x){return subgroupShuffleXor(v,x);} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + AH2 AWaveXorH2(AH2 v,AU1 x){return AH2_AU1(subgroupShuffleXor(AU1_AH2(v),x));} + AH4 AWaveXorH4(AH4 v,AU1 x){return AH4_AU2(subgroupShuffleXor(AU2_AH4(v),x));} + AW2 AWaveXorW2(AW2 v,AU1 x){return AW2_AU1(subgroupShuffleXor(AU1_AW2(v),x));} + AW4 AWaveXorW4(AW4 v,AU1 x){return AW4_AU2(subgroupShuffleXor(AU2_AW4(v),x));} + #endif + #endif +//============================================================================================================================== +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// HLSL +// +// +//============================================================================================================================== +#if defined(A_HLSL) && defined(A_GPU) + #ifdef A_HLSL_6_2 + #define AP1 bool + #define AP2 bool2 + #define AP3 bool3 + #define AP4 bool4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF1 float32_t + #define AF2 float32_t2 + #define AF3 float32_t3 + #define AF4 float32_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1 uint32_t + #define AU2 uint32_t2 + #define AU3 uint32_t3 + #define AU4 uint32_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASU1 int32_t + #define ASU2 int32_t2 + #define ASU3 int32_t3 + #define ASU4 int32_t4 + #else + #define AP1 bool + #define AP2 bool2 + #define AP3 bool3 + #define AP4 bool4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF1 float + #define AF2 float2 + #define AF3 float3 + #define AF4 float4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1 uint + #define AU2 uint2 + #define AU3 uint3 + #define AU4 uint4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASU1 int + #define ASU2 int2 + #define ASU3 int3 + #define ASU4 int4 + #endif +//============================================================================================================================== + #define AF1_AU1(x) asfloat(AU1(x)) + #define AF2_AU2(x) asfloat(AU2(x)) + #define AF3_AU3(x) asfloat(AU3(x)) + #define AF4_AU4(x) asfloat(AU4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AF1(x) asuint(AF1(x)) + #define AU2_AF2(x) asuint(AF2(x)) + #define AU3_AF3(x) asuint(AF3(x)) + #define AU4_AF4(x) asuint(AF4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH1_AF1_x(AF1 a){return f32tof16(a);} + #define AU1_AH1_AF1(a) AU1_AH1_AF1_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH2_AF2_x(AF2 a){return f32tof16(a.x)|(f32tof16(a.y)<<16);} + #define AU1_AH2_AF2(a) AU1_AH2_AF2_x(AF2(a)) + #define AU1_AB4Unorm_AF4(x) D3DCOLORtoUBYTE4(AF4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AF2 AF2_AH2_AU1_x(AU1 x){return AF2(f16tof32(x&0xFFFF),f16tof32(x>>16));} + #define AF2_AH2_AU1(x) AF2_AH2_AU1_x(AU1(x)) +//============================================================================================================================== + AF1 AF1_x(AF1 a){return AF1(a);} + AF2 AF2_x(AF1 a){return AF2(a,a);} + AF3 AF3_x(AF1 a){return AF3(a,a,a);} + AF4 AF4_x(AF1 a){return AF4(a,a,a,a);} + #define AF1_(a) AF1_x(AF1(a)) + #define AF2_(a) AF2_x(AF1(a)) + #define AF3_(a) AF3_x(AF1(a)) + #define AF4_(a) AF4_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_x(AU1 a){return AU1(a);} + AU2 AU2_x(AU1 a){return AU2(a,a);} + AU3 AU3_x(AU1 a){return AU3(a,a,a);} + AU4 AU4_x(AU1 a){return AU4(a,a,a,a);} + #define AU1_(a) AU1_x(AU1(a)) + #define AU2_(a) AU2_x(AU1(a)) + #define AU3_(a) AU3_x(AU1(a)) + #define AU4_(a) AU4_x(AU1(a)) +//============================================================================================================================== + AU1 AAbsSU1(AU1 a){return AU1(abs(ASU1(a)));} + AU2 AAbsSU2(AU2 a){return AU2(abs(ASU2(a)));} + AU3 AAbsSU3(AU3 a){return AU3(abs(ASU3(a)));} + AU4 AAbsSU4(AU4 a){return AU4(abs(ASU4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 ABfe(AU1 src,AU1 off,AU1 bits){AU1 mask=(1u<>off)&mask;} + AU1 ABfi(AU1 src,AU1 ins,AU1 mask){return (ins&mask)|(src&(~mask));} + AU1 ABfiM(AU1 src,AU1 ins,AU1 bits){AU1 mask=(1u<>ASU1(b));} + AU2 AShrSU2(AU2 a,AU2 b){return AU2(ASU2(a)>>ASU2(b));} + AU3 AShrSU3(AU3 a,AU3 b){return AU3(ASU3(a)>>ASU3(b));} + AU4 AShrSU4(AU4 a,AU4 b){return AU4(ASU4(a)>>ASU4(b));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HLSL BYTE +//============================================================================================================================== + #ifdef A_BYTE + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HLSL HALF +//============================================================================================================================== + #ifdef A_HALF + #ifdef A_HLSL_6_2 + #define AH1 float16_t + #define AH2 float16_t2 + #define AH3 float16_t3 + #define AH4 float16_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AW1 uint16_t + #define AW2 uint16_t2 + #define AW3 uint16_t3 + #define AW4 uint16_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASW1 int16_t + #define ASW2 int16_t2 + #define ASW3 int16_t3 + #define ASW4 int16_t4 + #else + #define AH1 min16float + #define AH2 min16float2 + #define AH3 min16float3 + #define AH4 min16float4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AW1 min16uint + #define AW2 min16uint2 + #define AW3 min16uint3 + #define AW4 min16uint4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASW1 min16int + #define ASW2 min16int2 + #define ASW3 min16int3 + #define ASW4 min16int4 + #endif +//============================================================================================================================== + // Need to use manual unpack to get optimal execution (don't use packed types in buffers directly). + // Unpack requires this pattern: https://gpuopen.com/first-steps-implementing-fp16/ + AH2 AH2_AU1_x(AU1 x){AF2 t=f16tof32(AU2(x&0xFFFF,x>>16));return AH2(t);} + AH4 AH4_AU2_x(AU2 x){return AH4(AH2_AU1_x(x.x),AH2_AU1_x(x.y));} + AW2 AW2_AU1_x(AU1 x){AU2 t=AU2(x&0xFFFF,x>>16);return AW2(t);} + AW4 AW4_AU2_x(AU2 x){return AW4(AW2_AU1_x(x.x),AW2_AU1_x(x.y));} + #define AH2_AU1(x) AH2_AU1_x(AU1(x)) + #define AH4_AU2(x) AH4_AU2_x(AU2(x)) + #define AW2_AU1(x) AW2_AU1_x(AU1(x)) + #define AW4_AU2(x) AW4_AU2_x(AU2(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH2_x(AH2 x){return f32tof16(x.x)+(f32tof16(x.y)<<16);} + AU2 AU2_AH4_x(AH4 x){return AU2(AU1_AH2_x(x.xy),AU1_AH2_x(x.zw));} + AU1 AU1_AW2_x(AW2 x){return AU1(x.x)+(AU1(x.y)<<16);} + AU2 AU2_AW4_x(AW4 x){return AU2(AU1_AW2_x(x.xy),AU1_AW2_x(x.zw));} + #define AU1_AH2(x) AU1_AH2_x(AH2(x)) + #define AU2_AH4(x) AU2_AH4_x(AH4(x)) + #define AU1_AW2(x) AU1_AW2_x(AW2(x)) + #define AU2_AW4(x) AU2_AW4_x(AW4(x)) +//============================================================================================================================== + #if defined(A_HLSL_6_2) && !defined(A_NO_16_BIT_CAST) + #define AW1_AH1(x) asuint16(x) + #define AW2_AH2(x) asuint16(x) + #define AW3_AH3(x) asuint16(x) + #define AW4_AH4(x) asuint16(x) + #else + #define AW1_AH1(a) AW1(f32tof16(AF1(a))) + #define AW2_AH2(a) AW2(AW1_AH1((a).x),AW1_AH1((a).y)) + #define AW3_AH3(a) AW3(AW1_AH1((a).x),AW1_AH1((a).y),AW1_AH1((a).z)) + #define AW4_AH4(a) AW4(AW1_AH1((a).x),AW1_AH1((a).y),AW1_AH1((a).z),AW1_AH1((a).w)) + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #if defined(A_HLSL_6_2) && !defined(A_NO_16_BIT_CAST) + #define AH1_AW1(x) asfloat16(x) + #define AH2_AW2(x) asfloat16(x) + #define AH3_AW3(x) asfloat16(x) + #define AH4_AW4(x) asfloat16(x) + #else + #define AH1_AW1(a) AH1(f16tof32(AU1(a))) + #define AH2_AW2(a) AH2(AH1_AW1((a).x),AH1_AW1((a).y)) + #define AH3_AW3(a) AH3(AH1_AW1((a).x),AH1_AW1((a).y),AH1_AW1((a).z)) + #define AH4_AW4(a) AH4(AH1_AW1((a).x),AH1_AW1((a).y),AH1_AW1((a).z),AH1_AW1((a).w)) + #endif +//============================================================================================================================== + AH1 AH1_x(AH1 a){return AH1(a);} + AH2 AH2_x(AH1 a){return AH2(a,a);} + AH3 AH3_x(AH1 a){return AH3(a,a,a);} + AH4 AH4_x(AH1 a){return AH4(a,a,a,a);} + #define AH1_(a) AH1_x(AH1(a)) + #define AH2_(a) AH2_x(AH1(a)) + #define AH3_(a) AH3_x(AH1(a)) + #define AH4_(a) AH4_x(AH1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AW1_x(AW1 a){return AW1(a);} + AW2 AW2_x(AW1 a){return AW2(a,a);} + AW3 AW3_x(AW1 a){return AW3(a,a,a);} + AW4 AW4_x(AW1 a){return AW4(a,a,a,a);} + #define AW1_(a) AW1_x(AW1(a)) + #define AW2_(a) AW2_x(AW1(a)) + #define AW3_(a) AW3_x(AW1(a)) + #define AW4_(a) AW4_x(AW1(a)) +//============================================================================================================================== + AW1 AAbsSW1(AW1 a){return AW1(abs(ASW1(a)));} + AW2 AAbsSW2(AW2 a){return AW2(abs(ASW2(a)));} + AW3 AAbsSW3(AW3 a){return AW3(abs(ASW3(a)));} + AW4 AAbsSW4(AW4 a){return AW4(abs(ASW4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AClampH1(AH1 x,AH1 n,AH1 m){return max(n,min(x,m));} + AH2 AClampH2(AH2 x,AH2 n,AH2 m){return max(n,min(x,m));} + AH3 AClampH3(AH3 x,AH3 n,AH3 m){return max(n,min(x,m));} + AH4 AClampH4(AH4 x,AH4 n,AH4 m){return max(n,min(x,m));} +//------------------------------------------------------------------------------------------------------------------------------ + // V_FRACT_F16 (note DX frac() is different). + AH1 AFractH1(AH1 x){return x-floor(x);} + AH2 AFractH2(AH2 x){return x-floor(x);} + AH3 AFractH3(AH3 x){return x-floor(x);} + AH4 AFractH4(AH4 x){return x-floor(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ALerpH1(AH1 x,AH1 y,AH1 a){return lerp(x,y,a);} + AH2 ALerpH2(AH2 x,AH2 y,AH2 a){return lerp(x,y,a);} + AH3 ALerpH3(AH3 x,AH3 y,AH3 a){return lerp(x,y,a);} + AH4 ALerpH4(AH4 x,AH4 y,AH4 a){return lerp(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AMax3H1(AH1 x,AH1 y,AH1 z){return max(x,max(y,z));} + AH2 AMax3H2(AH2 x,AH2 y,AH2 z){return max(x,max(y,z));} + AH3 AMax3H3(AH3 x,AH3 y,AH3 z){return max(x,max(y,z));} + AH4 AMax3H4(AH4 x,AH4 y,AH4 z){return max(x,max(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMaxSW1(AW1 a,AW1 b){return AW1(max(ASU1(a),ASU1(b)));} + AW2 AMaxSW2(AW2 a,AW2 b){return AW2(max(ASU2(a),ASU2(b)));} + AW3 AMaxSW3(AW3 a,AW3 b){return AW3(max(ASU3(a),ASU3(b)));} + AW4 AMaxSW4(AW4 a,AW4 b){return AW4(max(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AMin3H1(AH1 x,AH1 y,AH1 z){return min(x,min(y,z));} + AH2 AMin3H2(AH2 x,AH2 y,AH2 z){return min(x,min(y,z));} + AH3 AMin3H3(AH3 x,AH3 y,AH3 z){return min(x,min(y,z));} + AH4 AMin3H4(AH4 x,AH4 y,AH4 z){return min(x,min(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMinSW1(AW1 a,AW1 b){return AW1(min(ASU1(a),ASU1(b)));} + AW2 AMinSW2(AW2 a,AW2 b){return AW2(min(ASU2(a),ASU2(b)));} + AW3 AMinSW3(AW3 a,AW3 b){return AW3(min(ASU3(a),ASU3(b)));} + AW4 AMinSW4(AW4 a,AW4 b){return AW4(min(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARcpH1(AH1 x){return rcp(x);} + AH2 ARcpH2(AH2 x){return rcp(x);} + AH3 ARcpH3(AH3 x){return rcp(x);} + AH4 ARcpH4(AH4 x){return rcp(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARsqH1(AH1 x){return rsqrt(x);} + AH2 ARsqH2(AH2 x){return rsqrt(x);} + AH3 ARsqH3(AH3 x){return rsqrt(x);} + AH4 ARsqH4(AH4 x){return rsqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASatH1(AH1 x){return saturate(x);} + AH2 ASatH2(AH2 x){return saturate(x);} + AH3 ASatH3(AH3 x){return saturate(x);} + AH4 ASatH4(AH4 x){return saturate(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AShrSW1(AW1 a,AW1 b){return AW1(ASW1(a)>>ASW1(b));} + AW2 AShrSW2(AW2 a,AW2 b){return AW2(ASW2(a)>>ASW2(b));} + AW3 AShrSW3(AW3 a,AW3 b){return AW3(ASW3(a)>>ASW3(b));} + AW4 AShrSW4(AW4 a,AW4 b){return AW4(ASW4(a)>>ASW4(b));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HLSL DOUBLE +//============================================================================================================================== + #ifdef A_DUBL + #ifdef A_HLSL_6_2 + #define AD1 float64_t + #define AD2 float64_t2 + #define AD3 float64_t3 + #define AD4 float64_t4 + #else + #define AD1 double + #define AD2 double2 + #define AD3 double3 + #define AD4 double4 + #endif +//------------------------------------------------------------------------------------------------------------------------------ + AD1 AD1_x(AD1 a){return AD1(a);} + AD2 AD2_x(AD1 a){return AD2(a,a);} + AD3 AD3_x(AD1 a){return AD3(a,a,a);} + AD4 AD4_x(AD1 a){return AD4(a,a,a,a);} + #define AD1_(a) AD1_x(AD1(a)) + #define AD2_(a) AD2_x(AD1(a)) + #define AD3_(a) AD3_x(AD1(a)) + #define AD4_(a) AD4_x(AD1(a)) +//============================================================================================================================== + AD1 AFractD1(AD1 a){return a-floor(a);} + AD2 AFractD2(AD2 a){return a-floor(a);} + AD3 AFractD3(AD3 a){return a-floor(a);} + AD4 AFractD4(AD4 a){return a-floor(a);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ALerpD1(AD1 x,AD1 y,AD1 a){return lerp(x,y,a);} + AD2 ALerpD2(AD2 x,AD2 y,AD2 a){return lerp(x,y,a);} + AD3 ALerpD3(AD3 x,AD3 y,AD3 a){return lerp(x,y,a);} + AD4 ALerpD4(AD4 x,AD4 y,AD4 a){return lerp(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARcpD1(AD1 x){return rcp(x);} + AD2 ARcpD2(AD2 x){return rcp(x);} + AD3 ARcpD3(AD3 x){return rcp(x);} + AD4 ARcpD4(AD4 x){return rcp(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARsqD1(AD1 x){return rsqrt(x);} + AD2 ARsqD2(AD2 x){return rsqrt(x);} + AD3 ARsqD3(AD3 x){return rsqrt(x);} + AD4 ARsqD4(AD4 x){return rsqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ASatD1(AD1 x){return saturate(x);} + AD2 ASatD2(AD2 x){return saturate(x);} + AD3 ASatD3(AD3 x){return saturate(x);} + AD4 ASatD4(AD4 x){return saturate(x);} + #endif +//============================================================================================================================== +// HLSL WAVE +//============================================================================================================================== + #ifdef A_WAVE + // Where 'x' must be a compile time literal. + AF1 AWaveXorF1(AF1 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AF2 AWaveXorF2(AF2 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AF3 AWaveXorF3(AF3 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AF4 AWaveXorF4(AF4 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU1 AWaveXorU1(AU1 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU2 AWaveXorU1(AU2 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU3 AWaveXorU1(AU3 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU4 AWaveXorU1(AU4 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + AH2 AWaveXorH2(AH2 v,AU1 x){return AH2_AU1(WaveReadLaneAt(AU1_AH2(v),WaveGetLaneIndex()^x));} + AH4 AWaveXorH4(AH4 v,AU1 x){return AH4_AU2(WaveReadLaneAt(AU2_AH4(v),WaveGetLaneIndex()^x));} + AW2 AWaveXorW2(AW2 v,AU1 x){return AW2_AU1(WaveReadLaneAt(AU1_AW2(v),WaveGetLaneIndex()^x));} + AW4 AWaveXorW4(AW4 v,AU1 x){return AW4_AU1(WaveReadLaneAt(AU1_AW4(v),WaveGetLaneIndex()^x));} + #endif + #endif +//============================================================================================================================== +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// GPU COMMON +// +// +//============================================================================================================================== +#ifdef A_GPU + // Negative and positive infinity. + #define A_INFP_F AF1_AU1(0x7f800000u) + #define A_INFN_F AF1_AU1(0xff800000u) +//------------------------------------------------------------------------------------------------------------------------------ + // Copy sign from 's' to positive 'd'. + AF1 ACpySgnF1(AF1 d,AF1 s){return AF1_AU1(AU1_AF1(d)|(AU1_AF1(s)&AU1_(0x80000000u)));} + AF2 ACpySgnF2(AF2 d,AF2 s){return AF2_AU2(AU2_AF2(d)|(AU2_AF2(s)&AU2_(0x80000000u)));} + AF3 ACpySgnF3(AF3 d,AF3 s){return AF3_AU3(AU3_AF3(d)|(AU3_AF3(s)&AU3_(0x80000000u)));} + AF4 ACpySgnF4(AF4 d,AF4 s){return AF4_AU4(AU4_AF4(d)|(AU4_AF4(s)&AU4_(0x80000000u)));} +//------------------------------------------------------------------------------------------------------------------------------ + // Single operation to return (useful to create a mask to use in lerp for branch free logic), + // m=NaN := 0 + // m>=0 := 0 + // m<0 := 1 + // Uses the following useful floating point logic, + // saturate(+a*(-INF)==-INF) := 0 + // saturate( 0*(-INF)== NaN) := 0 + // saturate(-a*(-INF)==+INF) := 1 + AF1 ASignedF1(AF1 m){return ASatF1(m*AF1_(A_INFN_F));} + AF2 ASignedF2(AF2 m){return ASatF2(m*AF2_(A_INFN_F));} + AF3 ASignedF3(AF3 m){return ASatF3(m*AF3_(A_INFN_F));} + AF4 ASignedF4(AF4 m){return ASatF4(m*AF4_(A_INFN_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AGtZeroF1(AF1 m){return ASatF1(m*AF1_(A_INFP_F));} + AF2 AGtZeroF2(AF2 m){return ASatF2(m*AF2_(A_INFP_F));} + AF3 AGtZeroF3(AF3 m){return ASatF3(m*AF3_(A_INFP_F));} + AF4 AGtZeroF4(AF4 m){return ASatF4(m*AF4_(A_INFP_F));} +//============================================================================================================================== + #ifdef A_HALF + #ifdef A_HLSL_6_2 + #define A_INFP_H AH1_AW1((uint16_t)0x7c00u) + #define A_INFN_H AH1_AW1((uint16_t)0xfc00u) + #else + #define A_INFP_H AH1_AW1(0x7c00u) + #define A_INFN_H AH1_AW1(0xfc00u) + #endif + +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ACpySgnH1(AH1 d,AH1 s){return AH1_AW1(AW1_AH1(d)|(AW1_AH1(s)&AW1_(0x8000u)));} + AH2 ACpySgnH2(AH2 d,AH2 s){return AH2_AW2(AW2_AH2(d)|(AW2_AH2(s)&AW2_(0x8000u)));} + AH3 ACpySgnH3(AH3 d,AH3 s){return AH3_AW3(AW3_AH3(d)|(AW3_AH3(s)&AW3_(0x8000u)));} + AH4 ACpySgnH4(AH4 d,AH4 s){return AH4_AW4(AW4_AH4(d)|(AW4_AH4(s)&AW4_(0x8000u)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASignedH1(AH1 m){return ASatH1(m*AH1_(A_INFN_H));} + AH2 ASignedH2(AH2 m){return ASatH2(m*AH2_(A_INFN_H));} + AH3 ASignedH3(AH3 m){return ASatH3(m*AH3_(A_INFN_H));} + AH4 ASignedH4(AH4 m){return ASatH4(m*AH4_(A_INFN_H));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AGtZeroH1(AH1 m){return ASatH1(m*AH1_(A_INFP_H));} + AH2 AGtZeroH2(AH2 m){return ASatH2(m*AH2_(A_INFP_H));} + AH3 AGtZeroH3(AH3 m){return ASatH3(m*AH3_(A_INFP_H));} + AH4 AGtZeroH4(AH4 m){return ASatH4(m*AH4_(A_INFP_H));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [FIS] FLOAT INTEGER SORTABLE +//------------------------------------------------------------------------------------------------------------------------------ +// Float to integer sortable. +// - If sign bit=0, flip the sign bit (positives). +// - If sign bit=1, flip all bits (negatives). +// Integer sortable to float. +// - If sign bit=1, flip the sign bit (positives). +// - If sign bit=0, flip all bits (negatives). +// Has nice side effects. +// - Larger integers are more positive values. +// - Float zero is mapped to center of integers (so clear to integer zero is a nice default for atomic max usage). +// Burns 3 ops for conversion {shift,or,xor}. +//============================================================================================================================== + AU1 AFisToU1(AU1 x){return x^(( AShrSU1(x,AU1_(31)))|AU1_(0x80000000));} + AU1 AFisFromU1(AU1 x){return x^((~AShrSU1(x,AU1_(31)))|AU1_(0x80000000));} +//------------------------------------------------------------------------------------------------------------------------------ + // Just adjust high 16-bit value (useful when upper part of 32-bit word is a 16-bit float value). + AU1 AFisToHiU1(AU1 x){return x^(( AShrSU1(x,AU1_(15)))|AU1_(0x80000000));} + AU1 AFisFromHiU1(AU1 x){return x^((~AShrSU1(x,AU1_(15)))|AU1_(0x80000000));} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + AW1 AFisToW1(AW1 x){return x^(( AShrSW1(x,AW1_(15)))|AW1_(0x8000));} + AW1 AFisFromW1(AW1 x){return x^((~AShrSW1(x,AW1_(15)))|AW1_(0x8000));} +//------------------------------------------------------------------------------------------------------------------------------ + AW2 AFisToW2(AW2 x){return x^(( AShrSW2(x,AW2_(15)))|AW2_(0x8000));} + AW2 AFisFromW2(AW2 x){return x^((~AShrSW2(x,AW2_(15)))|AW2_(0x8000));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [PERM] V_PERM_B32 +//------------------------------------------------------------------------------------------------------------------------------ +// Support for V_PERM_B32 started in the 3rd generation of GCN. +//------------------------------------------------------------------------------------------------------------------------------ +// yyyyxxxx - The 'i' input. +// 76543210 +// ======== +// HGFEDCBA - Naming on permutation. +//------------------------------------------------------------------------------------------------------------------------------ +// TODO +// ==== +// - Make sure compiler optimizes this. +//============================================================================================================================== + #ifdef A_HALF + AU1 APerm0E0A(AU2 i){return((i.x )&0xffu)|((i.y<<16)&0xff0000u);} + AU1 APerm0F0B(AU2 i){return((i.x>> 8)&0xffu)|((i.y<< 8)&0xff0000u);} + AU1 APerm0G0C(AU2 i){return((i.x>>16)&0xffu)|((i.y )&0xff0000u);} + AU1 APerm0H0D(AU2 i){return((i.x>>24)&0xffu)|((i.y>> 8)&0xff0000u);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 APermHGFA(AU2 i){return((i.x )&0x000000ffu)|(i.y&0xffffff00u);} + AU1 APermHGFC(AU2 i){return((i.x>>16)&0x000000ffu)|(i.y&0xffffff00u);} + AU1 APermHGAE(AU2 i){return((i.x<< 8)&0x0000ff00u)|(i.y&0xffff00ffu);} + AU1 APermHGCE(AU2 i){return((i.x>> 8)&0x0000ff00u)|(i.y&0xffff00ffu);} + AU1 APermHAFE(AU2 i){return((i.x<<16)&0x00ff0000u)|(i.y&0xff00ffffu);} + AU1 APermHCFE(AU2 i){return((i.x )&0x00ff0000u)|(i.y&0xff00ffffu);} + AU1 APermAGFE(AU2 i){return((i.x<<24)&0xff000000u)|(i.y&0x00ffffffu);} + AU1 APermCGFE(AU2 i){return((i.x<< 8)&0xff000000u)|(i.y&0x00ffffffu);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 APermGCEA(AU2 i){return((i.x)&0x00ff00ffu)|((i.y<<8)&0xff00ff00u);} + AU1 APermGECA(AU2 i){return(((i.x)&0xffu)|((i.x>>8)&0xff00u)|((i.y<<16)&0xff0000u)|((i.y<<8)&0xff000000u));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [BUC] BYTE UNSIGNED CONVERSION +//------------------------------------------------------------------------------------------------------------------------------ +// Designed to use the optimal conversion, enables the scaling to possibly be factored into other computation. +// Works on a range of {0 to A_BUC_<32,16>}, for <32-bit, and 16-bit> respectively. +//------------------------------------------------------------------------------------------------------------------------------ +// OPCODE NOTES +// ============ +// GCN does not do UNORM or SNORM for bytes in opcodes. +// - V_CVT_F32_UBYTE{0,1,2,3} - Unsigned byte to float. +// - V_CVT_PKACC_U8_F32 - Float to unsigned byte (does bit-field insert into 32-bit integer). +// V_PERM_B32 does byte packing with ability to zero fill bytes as well. +// - Can pull out byte values from two sources, and zero fill upper 8-bits of packed hi and lo. +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABuc{0,1,2,3}{To,From}U1() - Designed for V_CVT_F32_UBYTE* and V_CVT_PKACCUM_U8_F32 ops. +// ==== ===== +// 0 : 0 +// 1 : 1 +// ... +// 255 : 255 +// : 256 (just outside the encoding range) +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABuc{0,1,2,3}{To,From}U2() - Designed for 16-bit denormal tricks and V_PERM_B32. +// ==== ===== +// 0 : 0 +// 1 : 1/512 +// 2 : 1/256 +// ... +// 64 : 1/8 +// 128 : 1/4 +// 255 : 255/512 +// : 1/2 (just outside the encoding range) +//------------------------------------------------------------------------------------------------------------------------------ +// OPTIMAL IMPLEMENTATIONS ON AMD ARCHITECTURES +// ============================================ +// r=ABuc0FromU1(i) +// V_CVT_F32_UBYTE0 r,i +// -------------------------------------------- +// r=ABuc0ToU1(d,i) +// V_CVT_PKACCUM_U8_F32 r,i,0,d +// -------------------------------------------- +// d=ABuc0FromU2(i) +// Where 'k0' is an SGPR with 0x0E0A +// Where 'k1' is an SGPR with {32768.0} packed into the lower 16-bits +// V_PERM_B32 d,i.x,i.y,k0 +// V_PK_FMA_F16 d,d,k1.x,0 +// -------------------------------------------- +// r=ABuc0ToU2(d,i) +// Where 'k0' is an SGPR with {1.0/32768.0} packed into the lower 16-bits +// Where 'k1' is an SGPR with 0x???? +// Where 'k2' is an SGPR with 0x???? +// V_PK_FMA_F16 i,i,k0.x,0 +// V_PERM_B32 r.x,i,i,k1 +// V_PERM_B32 r.y,i,i,k2 +//============================================================================================================================== + // Peak range for 32-bit and 16-bit operations. + #define A_BUC_32 (255.0) + #define A_BUC_16 (255.0/512.0) +//============================================================================================================================== + #if 1 + // Designed to be one V_CVT_PKACCUM_U8_F32. + // The extra min is required to pattern match to V_CVT_PKACCUM_U8_F32. + AU1 ABuc0ToU1(AU1 d,AF1 i){return (d&0xffffff00u)|((min(AU1(i),255u) )&(0x000000ffu));} + AU1 ABuc1ToU1(AU1 d,AF1 i){return (d&0xffff00ffu)|((min(AU1(i),255u)<< 8)&(0x0000ff00u));} + AU1 ABuc2ToU1(AU1 d,AF1 i){return (d&0xff00ffffu)|((min(AU1(i),255u)<<16)&(0x00ff0000u));} + AU1 ABuc3ToU1(AU1 d,AF1 i){return (d&0x00ffffffu)|((min(AU1(i),255u)<<24)&(0xff000000u));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed to be one V_CVT_F32_UBYTE*. + AF1 ABuc0FromU1(AU1 i){return AF1((i )&255u);} + AF1 ABuc1FromU1(AU1 i){return AF1((i>> 8)&255u);} + AF1 ABuc2FromU1(AU1 i){return AF1((i>>16)&255u);} + AF1 ABuc3FromU1(AU1 i){return AF1((i>>24)&255u);} + #endif +//============================================================================================================================== + #ifdef A_HALF + // Takes {x0,x1} and {y0,y1} and builds {{x0,y0},{x1,y1}}. + AW2 ABuc01ToW2(AH2 x,AH2 y){x*=AH2_(1.0/32768.0);y*=AH2_(1.0/32768.0); + return AW2_AU1(APermGCEA(AU2(AU1_AW2(AW2_AH2(x)),AU1_AW2(AW2_AH2(y)))));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed for 3 ops to do SOA to AOS and conversion. + AU2 ABuc0ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABuc1ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABuc2ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABuc3ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed for 2 ops to do both AOS to SOA, and conversion. + AH2 ABuc0FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)))*AH2_(32768.0);} + AH2 ABuc1FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)))*AH2_(32768.0);} + AH2 ABuc2FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)))*AH2_(32768.0);} + AH2 ABuc3FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)))*AH2_(32768.0);} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [BSC] BYTE SIGNED CONVERSION +//------------------------------------------------------------------------------------------------------------------------------ +// Similar to [BUC]. +// Works on a range of {-/+ A_BSC_<32,16>}, for <32-bit, and 16-bit> respectively. +//------------------------------------------------------------------------------------------------------------------------------ +// ENCODING (without zero-based encoding) +// ======== +// 0 = unused (can be used to mean something else) +// 1 = lowest value +// 128 = exact zero center (zero based encoding +// 255 = highest value +//------------------------------------------------------------------------------------------------------------------------------ +// Zero-based [Zb] flips the MSB bit of the byte (making 128 "exact zero" actually zero). +// This is useful if there is a desire for cleared values to decode as zero. +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABsc{0,1,2,3}{To,From}U2() - Designed for 16-bit denormal tricks and V_PERM_B32. +// ==== ===== +// 0 : -127/512 (unused) +// 1 : -126/512 +// 2 : -125/512 +// ... +// 128 : 0 +// ... +// 255 : 127/512 +// : 1/4 (just outside the encoding range) +//============================================================================================================================== + // Peak range for 32-bit and 16-bit operations. + #define A_BSC_32 (127.0) + #define A_BSC_16 (127.0/512.0) +//============================================================================================================================== + #if 1 + AU1 ABsc0ToU1(AU1 d,AF1 i){return (d&0xffffff00u)|((min(AU1(i+128.0),255u) )&(0x000000ffu));} + AU1 ABsc1ToU1(AU1 d,AF1 i){return (d&0xffff00ffu)|((min(AU1(i+128.0),255u)<< 8)&(0x0000ff00u));} + AU1 ABsc2ToU1(AU1 d,AF1 i){return (d&0xff00ffffu)|((min(AU1(i+128.0),255u)<<16)&(0x00ff0000u));} + AU1 ABsc3ToU1(AU1 d,AF1 i){return (d&0x00ffffffu)|((min(AU1(i+128.0),255u)<<24)&(0xff000000u));} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 ABsc0ToZbU1(AU1 d,AF1 i){return ((d&0xffffff00u)|((min(AU1(trunc(i)+128.0),255u) )&(0x000000ffu)))^0x00000080u;} + AU1 ABsc1ToZbU1(AU1 d,AF1 i){return ((d&0xffff00ffu)|((min(AU1(trunc(i)+128.0),255u)<< 8)&(0x0000ff00u)))^0x00008000u;} + AU1 ABsc2ToZbU1(AU1 d,AF1 i){return ((d&0xff00ffffu)|((min(AU1(trunc(i)+128.0),255u)<<16)&(0x00ff0000u)))^0x00800000u;} + AU1 ABsc3ToZbU1(AU1 d,AF1 i){return ((d&0x00ffffffu)|((min(AU1(trunc(i)+128.0),255u)<<24)&(0xff000000u)))^0x80000000u;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ABsc0FromU1(AU1 i){return AF1((i )&255u)-128.0;} + AF1 ABsc1FromU1(AU1 i){return AF1((i>> 8)&255u)-128.0;} + AF1 ABsc2FromU1(AU1 i){return AF1((i>>16)&255u)-128.0;} + AF1 ABsc3FromU1(AU1 i){return AF1((i>>24)&255u)-128.0;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ABsc0FromZbU1(AU1 i){return AF1(((i )&255u)^0x80u)-128.0;} + AF1 ABsc1FromZbU1(AU1 i){return AF1(((i>> 8)&255u)^0x80u)-128.0;} + AF1 ABsc2FromZbU1(AU1 i){return AF1(((i>>16)&255u)^0x80u)-128.0;} + AF1 ABsc3FromZbU1(AU1 i){return AF1(((i>>24)&255u)^0x80u)-128.0;} + #endif +//============================================================================================================================== + #ifdef A_HALF + // Takes {x0,x1} and {y0,y1} and builds {{x0,y0},{x1,y1}}. + AW2 ABsc01ToW2(AH2 x,AH2 y){x=x*AH2_(1.0/32768.0)+AH2_(0.25/32768.0);y=y*AH2_(1.0/32768.0)+AH2_(0.25/32768.0); + return AW2_AU1(APermGCEA(AU2(AU1_AW2(AW2_AH2(x)),AU1_AW2(AW2_AH2(y)))));} +//------------------------------------------------------------------------------------------------------------------------------ + AU2 ABsc0ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABsc1ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABsc2ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABsc3ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AU2 ABsc0ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABsc1ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABsc2ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABsc3ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH2 ABsc0FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc1FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc2FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc3FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)))*AH2_(32768.0)-AH2_(0.25);} +//------------------------------------------------------------------------------------------------------------------------------ + AH2 ABsc0FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc1FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc2FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc3FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HALF APPROXIMATIONS +//------------------------------------------------------------------------------------------------------------------------------ +// These support only positive inputs. +// Did not see value yet in specialization for range. +// Using quick testing, ended up mostly getting the same "best" approximation for various ranges. +// With hardware that can co-execute transcendentals, the value in approximations could be less than expected. +// However from a latency perspective, if execution of a transcendental is 4 clk, with no packed support, -> 8 clk total. +// And co-execution would require a compiler interleaving a lot of independent work for packed usage. +//------------------------------------------------------------------------------------------------------------------------------ +// The one Newton Raphson iteration form of rsq() was skipped (requires 6 ops total). +// Same with sqrt(), as this could be x*rsq() (7 ops). +//============================================================================================================================== + #ifdef A_HALF + // Minimize squared error across full positive range, 2 ops. + // The 0x1de2 based approximation maps {0 to 1} input maps to < 1 output. + AH1 APrxLoSqrtH1(AH1 a){return AH1_AW1((AW1_AH1(a)>>AW1_(1))+AW1_(0x1de2));} + AH2 APrxLoSqrtH2(AH2 a){return AH2_AW2((AW2_AH2(a)>>AW2_(1))+AW2_(0x1de2));} + AH3 APrxLoSqrtH3(AH3 a){return AH3_AW3((AW3_AH3(a)>>AW3_(1))+AW3_(0x1de2));} + AH4 APrxLoSqrtH4(AH4 a){return AH4_AW4((AW4_AH4(a)>>AW4_(1))+AW4_(0x1de2));} +//------------------------------------------------------------------------------------------------------------------------------ + // Lower precision estimation, 1 op. + // Minimize squared error across {smallest normal to 16384.0}. + AH1 APrxLoRcpH1(AH1 a){return AH1_AW1(AW1_(0x7784)-AW1_AH1(a));} + AH2 APrxLoRcpH2(AH2 a){return AH2_AW2(AW2_(0x7784)-AW2_AH2(a));} + AH3 APrxLoRcpH3(AH3 a){return AH3_AW3(AW3_(0x7784)-AW3_AH3(a));} + AH4 APrxLoRcpH4(AH4 a){return AH4_AW4(AW4_(0x7784)-AW4_AH4(a));} +//------------------------------------------------------------------------------------------------------------------------------ + // Medium precision estimation, one Newton Raphson iteration, 3 ops. + AH1 APrxMedRcpH1(AH1 a){AH1 b=AH1_AW1(AW1_(0x778d)-AW1_AH1(a));return b*(-b*a+AH1_(2.0));} + AH2 APrxMedRcpH2(AH2 a){AH2 b=AH2_AW2(AW2_(0x778d)-AW2_AH2(a));return b*(-b*a+AH2_(2.0));} + AH3 APrxMedRcpH3(AH3 a){AH3 b=AH3_AW3(AW3_(0x778d)-AW3_AH3(a));return b*(-b*a+AH3_(2.0));} + AH4 APrxMedRcpH4(AH4 a){AH4 b=AH4_AW4(AW4_(0x778d)-AW4_AH4(a));return b*(-b*a+AH4_(2.0));} +//------------------------------------------------------------------------------------------------------------------------------ + // Minimize squared error across {smallest normal to 16384.0}, 2 ops. + AH1 APrxLoRsqH1(AH1 a){return AH1_AW1(AW1_(0x59a3)-(AW1_AH1(a)>>AW1_(1)));} + AH2 APrxLoRsqH2(AH2 a){return AH2_AW2(AW2_(0x59a3)-(AW2_AH2(a)>>AW2_(1)));} + AH3 APrxLoRsqH3(AH3 a){return AH3_AW3(AW3_(0x59a3)-(AW3_AH3(a)>>AW3_(1)));} + AH4 APrxLoRsqH4(AH4 a){return AH4_AW4(AW4_(0x59a3)-(AW4_AH4(a)>>AW4_(1)));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// FLOAT APPROXIMATIONS +//------------------------------------------------------------------------------------------------------------------------------ +// Michal Drobot has an excellent presentation on these: "Low Level Optimizations For GCN", +// - Idea dates back to SGI, then to Quake 3, etc. +// - https://michaldrobot.files.wordpress.com/2014/05/gcn_alu_opt_digitaldragons2014.pdf +// - sqrt(x)=rsqrt(x)*x +// - rcp(x)=rsqrt(x)*rsqrt(x) for positive x +// - https://github.com/michaldrobot/ShaderFastLibs/blob/master/ShaderFastMathLib.h +//------------------------------------------------------------------------------------------------------------------------------ +// These below are from perhaps less complete searching for optimal. +// Used FP16 normal range for testing with +4096 32-bit step size for sampling error. +// So these match up well with the half approximations. +//============================================================================================================================== + AF1 APrxLoSqrtF1(AF1 a){return AF1_AU1((AU1_AF1(a)>>AU1_(1))+AU1_(0x1fbc4639));} + AF1 APrxLoRcpF1(AF1 a){return AF1_AU1(AU1_(0x7ef07ebb)-AU1_AF1(a));} + AF1 APrxMedRcpF1(AF1 a){AF1 b=AF1_AU1(AU1_(0x7ef19fff)-AU1_AF1(a));return b*(-b*a+AF1_(2.0));} + AF1 APrxLoRsqF1(AF1 a){return AF1_AU1(AU1_(0x5f347d74)-(AU1_AF1(a)>>AU1_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 APrxLoSqrtF2(AF2 a){return AF2_AU2((AU2_AF2(a)>>AU2_(1))+AU2_(0x1fbc4639));} + AF2 APrxLoRcpF2(AF2 a){return AF2_AU2(AU2_(0x7ef07ebb)-AU2_AF2(a));} + AF2 APrxMedRcpF2(AF2 a){AF2 b=AF2_AU2(AU2_(0x7ef19fff)-AU2_AF2(a));return b*(-b*a+AF2_(2.0));} + AF2 APrxLoRsqF2(AF2 a){return AF2_AU2(AU2_(0x5f347d74)-(AU2_AF2(a)>>AU2_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF3 APrxLoSqrtF3(AF3 a){return AF3_AU3((AU3_AF3(a)>>AU3_(1))+AU3_(0x1fbc4639));} + AF3 APrxLoRcpF3(AF3 a){return AF3_AU3(AU3_(0x7ef07ebb)-AU3_AF3(a));} + AF3 APrxMedRcpF3(AF3 a){AF3 b=AF3_AU3(AU3_(0x7ef19fff)-AU3_AF3(a));return b*(-b*a+AF3_(2.0));} + AF3 APrxLoRsqF3(AF3 a){return AF3_AU3(AU3_(0x5f347d74)-(AU3_AF3(a)>>AU3_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF4 APrxLoSqrtF4(AF4 a){return AF4_AU4((AU4_AF4(a)>>AU4_(1))+AU4_(0x1fbc4639));} + AF4 APrxLoRcpF4(AF4 a){return AF4_AU4(AU4_(0x7ef07ebb)-AU4_AF4(a));} + AF4 APrxMedRcpF4(AF4 a){AF4 b=AF4_AU4(AU4_(0x7ef19fff)-AU4_AF4(a));return b*(-b*a+AF4_(2.0));} + AF4 APrxLoRsqF4(AF4 a){return AF4_AU4(AU4_(0x5f347d74)-(AU4_AF4(a)>>AU4_(1)));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PQ APPROXIMATIONS +//------------------------------------------------------------------------------------------------------------------------------ +// PQ is very close to x^(1/8). The functions below Use the fast float approximation method to do +// PQ<~>Gamma2 (4th power and fast 4th root) and PQ<~>Linear (8th power and fast 8th root). Maximum error is ~0.2%. +//============================================================================================================================== +// Helpers + AF1 Quart(AF1 a) { a = a * a; return a * a;} + AF1 Oct(AF1 a) { a = a * a; a = a * a; return a * a; } + AF2 Quart(AF2 a) { a = a * a; return a * a; } + AF2 Oct(AF2 a) { a = a * a; a = a * a; return a * a; } + AF3 Quart(AF3 a) { a = a * a; return a * a; } + AF3 Oct(AF3 a) { a = a * a; a = a * a; return a * a; } + AF4 Quart(AF4 a) { a = a * a; return a * a; } + AF4 Oct(AF4 a) { a = a * a; a = a * a; return a * a; } + //------------------------------------------------------------------------------------------------------------------------------ + AF1 APrxPQToGamma2(AF1 a) { return Quart(a); } + AF1 APrxPQToLinear(AF1 a) { return Oct(a); } + AF1 APrxLoGamma2ToPQ(AF1 a) { return AF1_AU1((AU1_AF1(a) >> AU1_(2)) + AU1_(0x2F9A4E46)); } + AF1 APrxMedGamma2ToPQ(AF1 a) { AF1 b = AF1_AU1((AU1_AF1(a) >> AU1_(2)) + AU1_(0x2F9A4E46)); AF1 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF1 APrxHighGamma2ToPQ(AF1 a) { return sqrt(sqrt(a)); } + AF1 APrxLoLinearToPQ(AF1 a) { return AF1_AU1((AU1_AF1(a) >> AU1_(3)) + AU1_(0x378D8723)); } + AF1 APrxMedLinearToPQ(AF1 a) { AF1 b = AF1_AU1((AU1_AF1(a) >> AU1_(3)) + AU1_(0x378D8723)); AF1 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF1 APrxHighLinearToPQ(AF1 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF2 APrxPQToGamma2(AF2 a) { return Quart(a); } + AF2 APrxPQToLinear(AF2 a) { return Oct(a); } + AF2 APrxLoGamma2ToPQ(AF2 a) { return AF2_AU2((AU2_AF2(a) >> AU2_(2)) + AU2_(0x2F9A4E46)); } + AF2 APrxMedGamma2ToPQ(AF2 a) { AF2 b = AF2_AU2((AU2_AF2(a) >> AU2_(2)) + AU2_(0x2F9A4E46)); AF2 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF2 APrxHighGamma2ToPQ(AF2 a) { return sqrt(sqrt(a)); } + AF2 APrxLoLinearToPQ(AF2 a) { return AF2_AU2((AU2_AF2(a) >> AU2_(3)) + AU2_(0x378D8723)); } + AF2 APrxMedLinearToPQ(AF2 a) { AF2 b = AF2_AU2((AU2_AF2(a) >> AU2_(3)) + AU2_(0x378D8723)); AF2 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF2 APrxHighLinearToPQ(AF2 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF3 APrxPQToGamma2(AF3 a) { return Quart(a); } + AF3 APrxPQToLinear(AF3 a) { return Oct(a); } + AF3 APrxLoGamma2ToPQ(AF3 a) { return AF3_AU3((AU3_AF3(a) >> AU3_(2)) + AU3_(0x2F9A4E46)); } + AF3 APrxMedGamma2ToPQ(AF3 a) { AF3 b = AF3_AU3((AU3_AF3(a) >> AU3_(2)) + AU3_(0x2F9A4E46)); AF3 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF3 APrxHighGamma2ToPQ(AF3 a) { return sqrt(sqrt(a)); } + AF3 APrxLoLinearToPQ(AF3 a) { return AF3_AU3((AU3_AF3(a) >> AU3_(3)) + AU3_(0x378D8723)); } + AF3 APrxMedLinearToPQ(AF3 a) { AF3 b = AF3_AU3((AU3_AF3(a) >> AU3_(3)) + AU3_(0x378D8723)); AF3 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF3 APrxHighLinearToPQ(AF3 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF4 APrxPQToGamma2(AF4 a) { return Quart(a); } + AF4 APrxPQToLinear(AF4 a) { return Oct(a); } + AF4 APrxLoGamma2ToPQ(AF4 a) { return AF4_AU4((AU4_AF4(a) >> AU4_(2)) + AU4_(0x2F9A4E46)); } + AF4 APrxMedGamma2ToPQ(AF4 a) { AF4 b = AF4_AU4((AU4_AF4(a) >> AU4_(2)) + AU4_(0x2F9A4E46)); AF4 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF4 APrxHighGamma2ToPQ(AF4 a) { return sqrt(sqrt(a)); } + AF4 APrxLoLinearToPQ(AF4 a) { return AF4_AU4((AU4_AF4(a) >> AU4_(3)) + AU4_(0x378D8723)); } + AF4 APrxMedLinearToPQ(AF4 a) { AF4 b = AF4_AU4((AU4_AF4(a) >> AU4_(3)) + AU4_(0x378D8723)); AF4 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF4 APrxHighLinearToPQ(AF4 a) { return sqrt(sqrt(sqrt(a))); } +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PARABOLIC SIN & COS +//------------------------------------------------------------------------------------------------------------------------------ +// Approximate answers to transcendental questions. +//------------------------------------------------------------------------------------------------------------------------------ +//============================================================================================================================== + #if 1 + // Valid input range is {-1 to 1} representing {0 to 2 pi}. + // Output range is {-1/4 to 1/4} representing {-1 to 1}. + AF1 APSinF1(AF1 x){return x*abs(x)-x;} // MAD. + AF2 APSinF2(AF2 x){return x*abs(x)-x;} + AF1 APCosF1(AF1 x){x=AFractF1(x*AF1_(0.5)+AF1_(0.75));x=x*AF1_(2.0)-AF1_(1.0);return APSinF1(x);} // 3x MAD, FRACT + AF2 APCosF2(AF2 x){x=AFractF2(x*AF2_(0.5)+AF2_(0.75));x=x*AF2_(2.0)-AF2_(1.0);return APSinF2(x);} + AF2 APSinCosF1(AF1 x){AF1 y=AFractF1(x*AF1_(0.5)+AF1_(0.75));y=y*AF1_(2.0)-AF1_(1.0);return APSinF2(AF2(x,y));} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + // For a packed {sin,cos} pair, + // - Native takes 16 clocks and 4 issue slots (no packed transcendentals). + // - Parabolic takes 8 clocks and 8 issue slots (only fract is non-packed). + AH1 APSinH1(AH1 x){return x*abs(x)-x;} + AH2 APSinH2(AH2 x){return x*abs(x)-x;} // AND,FMA + AH1 APCosH1(AH1 x){x=AFractH1(x*AH1_(0.5)+AH1_(0.75));x=x*AH1_(2.0)-AH1_(1.0);return APSinH1(x);} + AH2 APCosH2(AH2 x){x=AFractH2(x*AH2_(0.5)+AH2_(0.75));x=x*AH2_(2.0)-AH2_(1.0);return APSinH2(x);} // 3x FMA, 2xFRACT, AND + AH2 APSinCosH1(AH1 x){AH1 y=AFractH1(x*AH1_(0.5)+AH1_(0.75));y=y*AH1_(2.0)-AH1_(1.0);return APSinH2(AH2(x,y));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [ZOL] ZERO ONE LOGIC +//------------------------------------------------------------------------------------------------------------------------------ +// Conditional free logic designed for easy 16-bit packing, and backwards porting to 32-bit. +//------------------------------------------------------------------------------------------------------------------------------ +// 0 := false +// 1 := true +//------------------------------------------------------------------------------------------------------------------------------ +// AndNot(x,y) -> !(x&y) .... One op. +// AndOr(x,y,z) -> (x&y)|z ... One op. +// GtZero(x) -> x>0.0 ..... One op. +// Sel(x,y,z) -> x?y:z ..... Two ops, has no precision loss. +// Signed(x) -> x<0.0 ..... One op. +// ZeroPass(x,y) -> x?0:y ..... Two ops, 'y' is a pass through safe for aliasing as integer. +//------------------------------------------------------------------------------------------------------------------------------ +// OPTIMIZATION NOTES +// ================== +// - On Vega to use 2 constants in a packed op, pass in as one AW2 or one AH2 'k.xy' and use as 'k.xx' and 'k.yy'. +// For example 'a.xy*k.xx+k.yy'. +//============================================================================================================================== + #if 1 + AU1 AZolAndU1(AU1 x,AU1 y){return min(x,y);} + AU2 AZolAndU2(AU2 x,AU2 y){return min(x,y);} + AU3 AZolAndU3(AU3 x,AU3 y){return min(x,y);} + AU4 AZolAndU4(AU4 x,AU4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AZolNotU1(AU1 x){return x^AU1_(1);} + AU2 AZolNotU2(AU2 x){return x^AU2_(1);} + AU3 AZolNotU3(AU3 x){return x^AU3_(1);} + AU4 AZolNotU4(AU4 x){return x^AU4_(1);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AZolOrU1(AU1 x,AU1 y){return max(x,y);} + AU2 AZolOrU2(AU2 x,AU2 y){return max(x,y);} + AU3 AZolOrU3(AU3 x,AU3 y){return max(x,y);} + AU4 AZolOrU4(AU4 x,AU4 y){return max(x,y);} +//============================================================================================================================== + AU1 AZolF1ToU1(AF1 x){return AU1(x);} + AU2 AZolF2ToU2(AF2 x){return AU2(x);} + AU3 AZolF3ToU3(AF3 x){return AU3(x);} + AU4 AZolF4ToU4(AF4 x){return AU4(x);} +//------------------------------------------------------------------------------------------------------------------------------ + // 2 ops, denormals don't work in 32-bit on PC (and if they are enabled, OMOD is disabled). + AU1 AZolNotF1ToU1(AF1 x){return AU1(AF1_(1.0)-x);} + AU2 AZolNotF2ToU2(AF2 x){return AU2(AF2_(1.0)-x);} + AU3 AZolNotF3ToU3(AF3 x){return AU3(AF3_(1.0)-x);} + AU4 AZolNotF4ToU4(AF4 x){return AU4(AF4_(1.0)-x);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolU1ToF1(AU1 x){return AF1(x);} + AF2 AZolU2ToF2(AU2 x){return AF2(x);} + AF3 AZolU3ToF3(AU3 x){return AF3(x);} + AF4 AZolU4ToF4(AU4 x){return AF4(x);} +//============================================================================================================================== + AF1 AZolAndF1(AF1 x,AF1 y){return min(x,y);} + AF2 AZolAndF2(AF2 x,AF2 y){return min(x,y);} + AF3 AZolAndF3(AF3 x,AF3 y){return min(x,y);} + AF4 AZolAndF4(AF4 x,AF4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ASolAndNotF1(AF1 x,AF1 y){return (-x)*y+AF1_(1.0);} + AF2 ASolAndNotF2(AF2 x,AF2 y){return (-x)*y+AF2_(1.0);} + AF3 ASolAndNotF3(AF3 x,AF3 y){return (-x)*y+AF3_(1.0);} + AF4 ASolAndNotF4(AF4 x,AF4 y){return (-x)*y+AF4_(1.0);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolAndOrF1(AF1 x,AF1 y,AF1 z){return ASatF1(x*y+z);} + AF2 AZolAndOrF2(AF2 x,AF2 y,AF2 z){return ASatF2(x*y+z);} + AF3 AZolAndOrF3(AF3 x,AF3 y,AF3 z){return ASatF3(x*y+z);} + AF4 AZolAndOrF4(AF4 x,AF4 y,AF4 z){return ASatF4(x*y+z);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolGtZeroF1(AF1 x){return ASatF1(x*AF1_(A_INFP_F));} + AF2 AZolGtZeroF2(AF2 x){return ASatF2(x*AF2_(A_INFP_F));} + AF3 AZolGtZeroF3(AF3 x){return ASatF3(x*AF3_(A_INFP_F));} + AF4 AZolGtZeroF4(AF4 x){return ASatF4(x*AF4_(A_INFP_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolNotF1(AF1 x){return AF1_(1.0)-x;} + AF2 AZolNotF2(AF2 x){return AF2_(1.0)-x;} + AF3 AZolNotF3(AF3 x){return AF3_(1.0)-x;} + AF4 AZolNotF4(AF4 x){return AF4_(1.0)-x;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolOrF1(AF1 x,AF1 y){return max(x,y);} + AF2 AZolOrF2(AF2 x,AF2 y){return max(x,y);} + AF3 AZolOrF3(AF3 x,AF3 y){return max(x,y);} + AF4 AZolOrF4(AF4 x,AF4 y){return max(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolSelF1(AF1 x,AF1 y,AF1 z){AF1 r=(-x)*z+z;return x*y+r;} + AF2 AZolSelF2(AF2 x,AF2 y,AF2 z){AF2 r=(-x)*z+z;return x*y+r;} + AF3 AZolSelF3(AF3 x,AF3 y,AF3 z){AF3 r=(-x)*z+z;return x*y+r;} + AF4 AZolSelF4(AF4 x,AF4 y,AF4 z){AF4 r=(-x)*z+z;return x*y+r;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolSignedF1(AF1 x){return ASatF1(x*AF1_(A_INFN_F));} + AF2 AZolSignedF2(AF2 x){return ASatF2(x*AF2_(A_INFN_F));} + AF3 AZolSignedF3(AF3 x){return ASatF3(x*AF3_(A_INFN_F));} + AF4 AZolSignedF4(AF4 x){return ASatF4(x*AF4_(A_INFN_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolZeroPassF1(AF1 x,AF1 y){return AF1_AU1((AU1_AF1(x)!=AU1_(0))?AU1_(0):AU1_AF1(y));} + AF2 AZolZeroPassF2(AF2 x,AF2 y){return AF2_AU2((AU2_AF2(x)!=AU2_(0))?AU2_(0):AU2_AF2(y));} + AF3 AZolZeroPassF3(AF3 x,AF3 y){return AF3_AU3((AU3_AF3(x)!=AU3_(0))?AU3_(0):AU3_AF3(y));} + AF4 AZolZeroPassF4(AF4 x,AF4 y){return AF4_AU4((AU4_AF4(x)!=AU4_(0))?AU4_(0):AU4_AF4(y));} + #endif +//============================================================================================================================== + #ifdef A_HALF + AW1 AZolAndW1(AW1 x,AW1 y){return min(x,y);} + AW2 AZolAndW2(AW2 x,AW2 y){return min(x,y);} + AW3 AZolAndW3(AW3 x,AW3 y){return min(x,y);} + AW4 AZolAndW4(AW4 x,AW4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AZolNotW1(AW1 x){return x^AW1_(1);} + AW2 AZolNotW2(AW2 x){return x^AW2_(1);} + AW3 AZolNotW3(AW3 x){return x^AW3_(1);} + AW4 AZolNotW4(AW4 x){return x^AW4_(1);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AZolOrW1(AW1 x,AW1 y){return max(x,y);} + AW2 AZolOrW2(AW2 x,AW2 y){return max(x,y);} + AW3 AZolOrW3(AW3 x,AW3 y){return max(x,y);} + AW4 AZolOrW4(AW4 x,AW4 y){return max(x,y);} +//============================================================================================================================== + // Uses denormal trick. + AW1 AZolH1ToW1(AH1 x){return AW1_AH1(x*AH1_AW1(AW1_(1)));} + AW2 AZolH2ToW2(AH2 x){return AW2_AH2(x*AH2_AW2(AW2_(1)));} + AW3 AZolH3ToW3(AH3 x){return AW3_AH3(x*AH3_AW3(AW3_(1)));} + AW4 AZolH4ToW4(AH4 x){return AW4_AH4(x*AH4_AW4(AW4_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + // AMD arch lacks a packed conversion opcode. + AH1 AZolW1ToH1(AW1 x){return AH1_AW1(x*AW1_AH1(AH1_(1.0)));} + AH2 AZolW2ToH2(AW2 x){return AH2_AW2(x*AW2_AH2(AH2_(1.0)));} + AH3 AZolW1ToH3(AW3 x){return AH3_AW3(x*AW3_AH3(AH3_(1.0)));} + AH4 AZolW2ToH4(AW4 x){return AH4_AW4(x*AW4_AH4(AH4_(1.0)));} +//============================================================================================================================== + AH1 AZolAndH1(AH1 x,AH1 y){return min(x,y);} + AH2 AZolAndH2(AH2 x,AH2 y){return min(x,y);} + AH3 AZolAndH3(AH3 x,AH3 y){return min(x,y);} + AH4 AZolAndH4(AH4 x,AH4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASolAndNotH1(AH1 x,AH1 y){return (-x)*y+AH1_(1.0);} + AH2 ASolAndNotH2(AH2 x,AH2 y){return (-x)*y+AH2_(1.0);} + AH3 ASolAndNotH3(AH3 x,AH3 y){return (-x)*y+AH3_(1.0);} + AH4 ASolAndNotH4(AH4 x,AH4 y){return (-x)*y+AH4_(1.0);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolAndOrH1(AH1 x,AH1 y,AH1 z){return ASatH1(x*y+z);} + AH2 AZolAndOrH2(AH2 x,AH2 y,AH2 z){return ASatH2(x*y+z);} + AH3 AZolAndOrH3(AH3 x,AH3 y,AH3 z){return ASatH3(x*y+z);} + AH4 AZolAndOrH4(AH4 x,AH4 y,AH4 z){return ASatH4(x*y+z);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolGtZeroH1(AH1 x){return ASatH1(x*AH1_(A_INFP_H));} + AH2 AZolGtZeroH2(AH2 x){return ASatH2(x*AH2_(A_INFP_H));} + AH3 AZolGtZeroH3(AH3 x){return ASatH3(x*AH3_(A_INFP_H));} + AH4 AZolGtZeroH4(AH4 x){return ASatH4(x*AH4_(A_INFP_H));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolNotH1(AH1 x){return AH1_(1.0)-x;} + AH2 AZolNotH2(AH2 x){return AH2_(1.0)-x;} + AH3 AZolNotH3(AH3 x){return AH3_(1.0)-x;} + AH4 AZolNotH4(AH4 x){return AH4_(1.0)-x;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolOrH1(AH1 x,AH1 y){return max(x,y);} + AH2 AZolOrH2(AH2 x,AH2 y){return max(x,y);} + AH3 AZolOrH3(AH3 x,AH3 y){return max(x,y);} + AH4 AZolOrH4(AH4 x,AH4 y){return max(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolSelH1(AH1 x,AH1 y,AH1 z){AH1 r=(-x)*z+z;return x*y+r;} + AH2 AZolSelH2(AH2 x,AH2 y,AH2 z){AH2 r=(-x)*z+z;return x*y+r;} + AH3 AZolSelH3(AH3 x,AH3 y,AH3 z){AH3 r=(-x)*z+z;return x*y+r;} + AH4 AZolSelH4(AH4 x,AH4 y,AH4 z){AH4 r=(-x)*z+z;return x*y+r;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolSignedH1(AH1 x){return ASatH1(x*AH1_(A_INFN_H));} + AH2 AZolSignedH2(AH2 x){return ASatH2(x*AH2_(A_INFN_H));} + AH3 AZolSignedH3(AH3 x){return ASatH3(x*AH3_(A_INFN_H));} + AH4 AZolSignedH4(AH4 x){return ASatH4(x*AH4_(A_INFN_H));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// COLOR CONVERSIONS +//------------------------------------------------------------------------------------------------------------------------------ +// These are all linear to/from some other space (where 'linear' has been shortened out of the function name). +// So 'ToGamma' is 'LinearToGamma', and 'FromGamma' is 'LinearFromGamma'. +// These are branch free implementations. +// The AToSrgbF1() function is useful for stores for compute shaders for GPUs without hardware linear->sRGB store conversion. +//------------------------------------------------------------------------------------------------------------------------------ +// TRANSFER FUNCTIONS +// ================== +// 709 ..... Rec709 used for some HDTVs +// Gamma ... Typically 2.2 for some PC displays, or 2.4-2.5 for CRTs, or 2.2 FreeSync2 native +// Pq ...... PQ native for HDR10 +// Srgb .... The sRGB output, typical of PC displays, useful for 10-bit output, or storing to 8-bit UNORM without SRGB type +// Two ..... Gamma 2.0, fastest conversion (useful for intermediate pass approximations) +// Three ... Gamma 3.0, less fast, but good for HDR. +//------------------------------------------------------------------------------------------------------------------------------ +// KEEPING TO SPEC +// =============== +// Both Rec.709 and sRGB have a linear segment which as spec'ed would intersect the curved segment 2 times. +// (a.) For 8-bit sRGB, steps {0 to 10.3} are in the linear region (4% of the encoding range). +// (b.) For 8-bit 709, steps {0 to 20.7} are in the linear region (8% of the encoding range). +// Also there is a slight step in the transition regions. +// Precision of the coefficients in the spec being the likely cause. +// Main usage case of the sRGB code is to do the linear->sRGB converstion in a compute shader before store. +// This is to work around lack of hardware (typically only ROP does the conversion for free). +// To "correct" the linear segment, would be to introduce error, because hardware decode of sRGB->linear is fixed (and free). +// So this header keeps with the spec. +// For linear->sRGB transforms, the linear segment in some respects reduces error, because rounding in that region is linear. +// Rounding in the curved region in hardware (and fast software code) introduces error due to rounding in non-linear. +//------------------------------------------------------------------------------------------------------------------------------ +// FOR PQ +// ====== +// Both input and output is {0.0-1.0}, and where output 1.0 represents 10000.0 cd/m^2. +// All constants are only specified to FP32 precision. +// External PQ source reference, +// - https://github.com/ampas/aces-dev/blob/master/transforms/ctl/utilities/ACESlib.Utilities_Color.a1.0.1.ctl +//------------------------------------------------------------------------------------------------------------------------------ +// PACKED VERSIONS +// =============== +// These are the A*H2() functions. +// There is no PQ functions as FP16 seemed to not have enough precision for the conversion. +// The remaining functions are "good enough" for 8-bit, and maybe 10-bit if not concerned about a few 1-bit errors. +// Precision is lowest in the 709 conversion, higher in sRGB, higher still in Two and Gamma (when using 2.2 at least). +//------------------------------------------------------------------------------------------------------------------------------ +// NOTES +// ===== +// Could be faster for PQ conversions to be in ALU or a texture lookup depending on usage case. +//============================================================================================================================== + #if 1 + AF1 ATo709F1(AF1 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AF2 ATo709F2(AF2 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AF3 ATo709F3(AF3 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + // Note 'rcpX' is '1/x', where the 'x' is what would be used in AFromGamma(). + AF1 AToGammaF1(AF1 c,AF1 rcpX){return pow(c,AF1_(rcpX));} + AF2 AToGammaF2(AF2 c,AF1 rcpX){return pow(c,AF2_(rcpX));} + AF3 AToGammaF3(AF3 c,AF1 rcpX){return pow(c,AF3_(rcpX));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToPqF1(AF1 x){AF1 p=pow(x,AF1_(0.159302)); + return pow((AF1_(0.835938)+AF1_(18.8516)*p)/(AF1_(1.0)+AF1_(18.6875)*p),AF1_(78.8438));} + AF2 AToPqF1(AF2 x){AF2 p=pow(x,AF2_(0.159302)); + return pow((AF2_(0.835938)+AF2_(18.8516)*p)/(AF2_(1.0)+AF2_(18.6875)*p),AF2_(78.8438));} + AF3 AToPqF1(AF3 x){AF3 p=pow(x,AF3_(0.159302)); + return pow((AF3_(0.835938)+AF3_(18.8516)*p)/(AF3_(1.0)+AF3_(18.6875)*p),AF3_(78.8438));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToSrgbF1(AF1 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AF2 AToSrgbF2(AF2 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AF3 AToSrgbF3(AF3 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToTwoF1(AF1 c){return sqrt(c);} + AF2 AToTwoF2(AF2 c){return sqrt(c);} + AF3 AToTwoF3(AF3 c){return sqrt(c);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToThreeF1(AF1 c){return pow(c,AF1_(1.0/3.0));} + AF2 AToThreeF2(AF2 c){return pow(c,AF2_(1.0/3.0));} + AF3 AToThreeF3(AF3 c){return pow(c,AF3_(1.0/3.0));} + #endif +//============================================================================================================================== + #if 1 + // Unfortunately median won't work here. + AF1 AFrom709F1(AF1 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF1(AZolSignedF1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AF2 AFrom709F2(AF2 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF2(AZolSignedF2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AF3 AFrom709F3(AF3 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF3(AZolSignedF3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromGammaF1(AF1 c,AF1 x){return pow(c,AF1_(x));} + AF2 AFromGammaF2(AF2 c,AF1 x){return pow(c,AF2_(x));} + AF3 AFromGammaF3(AF3 c,AF1 x){return pow(c,AF3_(x));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromPqF1(AF1 x){AF1 p=pow(x,AF1_(0.0126833)); + return pow(ASatF1(p-AF1_(0.835938))/(AF1_(18.8516)-AF1_(18.6875)*p),AF1_(6.27739));} + AF2 AFromPqF1(AF2 x){AF2 p=pow(x,AF2_(0.0126833)); + return pow(ASatF2(p-AF2_(0.835938))/(AF2_(18.8516)-AF2_(18.6875)*p),AF2_(6.27739));} + AF3 AFromPqF1(AF3 x){AF3 p=pow(x,AF3_(0.0126833)); + return pow(ASatF3(p-AF3_(0.835938))/(AF3_(18.8516)-AF3_(18.6875)*p),AF3_(6.27739));} +//------------------------------------------------------------------------------------------------------------------------------ + // Unfortunately median won't work here. + AF1 AFromSrgbF1(AF1 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF1(AZolSignedF1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AF2 AFromSrgbF2(AF2 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF2(AZolSignedF2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AF3 AFromSrgbF3(AF3 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF3(AZolSignedF3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromTwoF1(AF1 c){return c*c;} + AF2 AFromTwoF2(AF2 c){return c*c;} + AF3 AFromTwoF3(AF3 c){return c*c;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromThreeF1(AF1 c){return c*c*c;} + AF2 AFromThreeF2(AF2 c){return c*c*c;} + AF3 AFromThreeF3(AF3 c){return c*c*c;} + #endif +//============================================================================================================================== + #ifdef A_HALF + AH1 ATo709H1(AH1 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AH2 ATo709H2(AH2 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AH3 ATo709H3(AH3 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToGammaH1(AH1 c,AH1 rcpX){return pow(c,AH1_(rcpX));} + AH2 AToGammaH2(AH2 c,AH1 rcpX){return pow(c,AH2_(rcpX));} + AH3 AToGammaH3(AH3 c,AH1 rcpX){return pow(c,AH3_(rcpX));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToSrgbH1(AH1 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AH2 AToSrgbH2(AH2 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AH3 AToSrgbH3(AH3 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToTwoH1(AH1 c){return sqrt(c);} + AH2 AToTwoH2(AH2 c){return sqrt(c);} + AH3 AToTwoH3(AH3 c){return sqrt(c);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToThreeF1(AH1 c){return pow(c,AH1_(1.0/3.0));} + AH2 AToThreeF2(AH2 c){return pow(c,AH2_(1.0/3.0));} + AH3 AToThreeF3(AH3 c){return pow(c,AH3_(1.0/3.0));} + #endif +//============================================================================================================================== + #ifdef A_HALF + AH1 AFrom709H1(AH1 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH1(AZolSignedH1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AH2 AFrom709H2(AH2 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH2(AZolSignedH2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AH3 AFrom709H3(AH3 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH3(AZolSignedH3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromGammaH1(AH1 c,AH1 x){return pow(c,AH1_(x));} + AH2 AFromGammaH2(AH2 c,AH1 x){return pow(c,AH2_(x));} + AH3 AFromGammaH3(AH3 c,AH1 x){return pow(c,AH3_(x));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AHromSrgbF1(AH1 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH1(AZolSignedH1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AH2 AHromSrgbF2(AH2 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH2(AZolSignedH2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AH3 AHromSrgbF3(AH3 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH3(AZolSignedH3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromTwoH1(AH1 c){return c*c;} + AH2 AFromTwoH2(AH2 c){return c*c;} + AH3 AFromTwoH3(AH3 c){return c*c;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromThreeH1(AH1 c){return c*c*c;} + AH2 AFromThreeH2(AH2 c){return c*c*c;} + AH3 AFromThreeH3(AH3 c){return c*c*c;} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CS REMAP +//============================================================================================================================== + // Simple remap 64x1 to 8x8 with rotated 2x2 pixel quads in quad linear. + // 543210 + // ====== + // ..xxx. + // yy...y + AU2 ARmp8x8(AU1 a){return AU2(ABfe(a,1u,3u),ABfiM(ABfe(a,3u,3u),a,1u));} +//============================================================================================================================== + // More complex remap 64x1 to 8x8 which is necessary for 2D wave reductions. + // 543210 + // ====== + // .xx..x + // y..yy. + // Details, + // LANE TO 8x8 MAPPING + // =================== + // 00 01 08 09 10 11 18 19 + // 02 03 0a 0b 12 13 1a 1b + // 04 05 0c 0d 14 15 1c 1d + // 06 07 0e 0f 16 17 1e 1f + // 20 21 28 29 30 31 38 39 + // 22 23 2a 2b 32 33 3a 3b + // 24 25 2c 2d 34 35 3c 3d + // 26 27 2e 2f 36 37 3e 3f + AU2 ARmpRed8x8(AU1 a){return AU2(ABfiM(ABfe(a,2u,3u),a,1u),ABfiM(ABfe(a,3u,3u),ABfe(a,1u,2u),2u));} +//============================================================================================================================== + #ifdef A_HALF + AW2 ARmp8x8H(AU1 a){return AW2(ABfe(a,1u,3u),ABfiM(ABfe(a,3u,3u),a,1u));} + AW2 ARmpRed8x8H(AU1 a){return AW2(ABfiM(ABfe(a,2u,3u),a,1u),ABfiM(ABfe(a,3u,3u),ABfe(a,1u,2u),2u));} + #endif +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// REFERENCE +// +//------------------------------------------------------------------------------------------------------------------------------ +// IEEE FLOAT RULES +// ================ +// - saturate(NaN)=0, saturate(-INF)=0, saturate(+INF)=1 +// - {+/-}0 * {+/-}INF = NaN +// - -INF + (+INF) = NaN +// - {+/-}0 / {+/-}0 = NaN +// - {+/-}INF / {+/-}INF = NaN +// - a<(-0) := sqrt(a) = NaN (a=-0.0 won't NaN) +// - 0 == -0 +// - 4/0 = +INF +// - 4/-0 = -INF +// - 4+INF = +INF +// - 4-INF = -INF +// - 4*(+INF) = +INF +// - 4*(-INF) = -INF +// - -4*(+INF) = -INF +// - sqrt(+INF) = +INF +//------------------------------------------------------------------------------------------------------------------------------ +// FP16 ENCODING +// ============= +// fedcba9876543210 +// ---------------- +// ......mmmmmmmmmm 10-bit mantissa (encodes 11-bit 0.5 to 1.0 except for denormals) +// .eeeee.......... 5-bit exponent +// .00000.......... denormals +// .00001.......... -14 exponent +// .11110.......... 15 exponent +// .111110000000000 infinity +// .11111nnnnnnnnnn NaN with n!=0 +// s............... sign +//------------------------------------------------------------------------------------------------------------------------------ +// FP16/INT16 ALIASING DENORMAL +// ============================ +// 11-bit unsigned integers alias with half float denormal/normal values, +// 1 = 2^(-24) = 1/16777216 ....................... first denormal value +// 2 = 2^(-23) +// ... +// 1023 = 2^(-14)*(1-2^(-10)) = 2^(-14)*(1-1/1024) ... last denormal value +// 1024 = 2^(-14) = 1/16384 .......................... first normal value that still maps to integers +// 2047 .............................................. last normal value that still maps to integers +// Scaling limits, +// 2^15 = 32768 ...................................... largest power of 2 scaling +// Largest pow2 conversion mapping is at *32768, +// 1 : 2^(-9) = 1/512 +// 2 : 1/256 +// 4 : 1/128 +// 8 : 1/64 +// 16 : 1/32 +// 32 : 1/16 +// 64 : 1/8 +// 128 : 1/4 +// 256 : 1/2 +// 512 : 1 +// 1024 : 2 +// 2047 : a little less than 4 +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// GPU/CPU PORTABILITY +// +// +//------------------------------------------------------------------------------------------------------------------------------ +// This is the GPU implementation. +// See the CPU implementation for docs. +//============================================================================================================================== +#ifdef A_GPU + #define A_TRUE true + #define A_FALSE false + #define A_STATIC +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR ARGUMENT/RETURN/INITIALIZATION PORTABILITY +//============================================================================================================================== + #define retAD2 AD2 + #define retAD3 AD3 + #define retAD4 AD4 + #define retAF2 AF2 + #define retAF3 AF3 + #define retAF4 AF4 + #define retAL2 AL2 + #define retAL3 AL3 + #define retAL4 AL4 + #define retAU2 AU2 + #define retAU3 AU3 + #define retAU4 AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define inAD2 in AD2 + #define inAD3 in AD3 + #define inAD4 in AD4 + #define inAF2 in AF2 + #define inAF3 in AF3 + #define inAF4 in AF4 + #define inAL2 in AL2 + #define inAL3 in AL3 + #define inAL4 in AL4 + #define inAU2 in AU2 + #define inAU3 in AU3 + #define inAU4 in AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define inoutAD2 inout AD2 + #define inoutAD3 inout AD3 + #define inoutAD4 inout AD4 + #define inoutAF2 inout AF2 + #define inoutAF3 inout AF3 + #define inoutAF4 inout AF4 + #define inoutAL2 inout AL2 + #define inoutAL3 inout AL3 + #define inoutAL4 inout AL4 + #define inoutAU2 inout AU2 + #define inoutAU3 inout AU3 + #define inoutAU4 inout AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define outAD2 out AD2 + #define outAD3 out AD3 + #define outAD4 out AD4 + #define outAF2 out AF2 + #define outAF3 out AF3 + #define outAF4 out AF4 + #define outAL2 out AL2 + #define outAL3 out AL3 + #define outAL4 out AL4 + #define outAU2 out AU2 + #define outAU3 out AU3 + #define outAU4 out AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define varAD2(x) AD2 x + #define varAD3(x) AD3 x + #define varAD4(x) AD4 x + #define varAF2(x) AF2 x + #define varAF3(x) AF3 x + #define varAF4(x) AF4 x + #define varAL2(x) AL2 x + #define varAL3(x) AL3 x + #define varAL4(x) AL4 x + #define varAU2(x) AU2 x + #define varAU3(x) AU3 x + #define varAU4(x) AU4 x +//------------------------------------------------------------------------------------------------------------------------------ + #define initAD2(x,y) AD2(x,y) + #define initAD3(x,y,z) AD3(x,y,z) + #define initAD4(x,y,z,w) AD4(x,y,z,w) + #define initAF2(x,y) AF2(x,y) + #define initAF3(x,y,z) AF3(x,y,z) + #define initAF4(x,y,z,w) AF4(x,y,z,w) + #define initAL2(x,y) AL2(x,y) + #define initAL3(x,y,z) AL3(x,y,z) + #define initAL4(x,y,z,w) AL4(x,y,z,w) + #define initAU2(x,y) AU2(x,y) + #define initAU3(x,y,z) AU3(x,y,z) + #define initAU4(x,y,z,w) AU4(x,y,z,w) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS +//============================================================================================================================== + #define AAbsD1(a) abs(AD1(a)) + #define AAbsF1(a) abs(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ACosD1(a) cos(AD1(a)) + #define ACosF1(a) cos(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ADotD2(a,b) dot(AD2(a),AD2(b)) + #define ADotD3(a,b) dot(AD3(a),AD3(b)) + #define ADotD4(a,b) dot(AD4(a),AD4(b)) + #define ADotF2(a,b) dot(AF2(a),AF2(b)) + #define ADotF3(a,b) dot(AF3(a),AF3(b)) + #define ADotF4(a,b) dot(AF4(a),AF4(b)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AExp2D1(a) exp2(AD1(a)) + #define AExp2F1(a) exp2(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AFloorD1(a) floor(AD1(a)) + #define AFloorF1(a) floor(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ALog2D1(a) log2(AD1(a)) + #define ALog2F1(a) log2(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AMaxD1(a,b) max(a,b) + #define AMaxF1(a,b) max(a,b) + #define AMaxL1(a,b) max(a,b) + #define AMaxU1(a,b) max(a,b) +//------------------------------------------------------------------------------------------------------------------------------ + #define AMinD1(a,b) min(a,b) + #define AMinF1(a,b) min(a,b) + #define AMinL1(a,b) min(a,b) + #define AMinU1(a,b) min(a,b) +//------------------------------------------------------------------------------------------------------------------------------ + #define ASinD1(a) sin(AD1(a)) + #define ASinF1(a) sin(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ASqrtD1(a) sqrt(AD1(a)) + #define ASqrtF1(a) sqrt(AF1(a)) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS - DEPENDENT +//============================================================================================================================== + #define APowD1(a,b) pow(AD1(a),AF1(b)) + #define APowF1(a,b) pow(AF1(a),AF1(b)) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR OPS +//------------------------------------------------------------------------------------------------------------------------------ +// These are added as needed for production or prototyping, so not necessarily a complete set. +// They follow a convention of taking in a destination and also returning the destination value to increase utility. +//============================================================================================================================== + #ifdef A_DUBL + AD2 opAAbsD2(outAD2 d,inAD2 a){d=abs(a);return d;} + AD3 opAAbsD3(outAD3 d,inAD3 a){d=abs(a);return d;} + AD4 opAAbsD4(outAD4 d,inAD4 a){d=abs(a);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAAddD2(outAD2 d,inAD2 a,inAD2 b){d=a+b;return d;} + AD3 opAAddD3(outAD3 d,inAD3 a,inAD3 b){d=a+b;return d;} + AD4 opAAddD4(outAD4 d,inAD4 a,inAD4 b){d=a+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAAddOneD2(outAD2 d,inAD2 a,AD1 b){d=a+AD2_(b);return d;} + AD3 opAAddOneD3(outAD3 d,inAD3 a,AD1 b){d=a+AD3_(b);return d;} + AD4 opAAddOneD4(outAD4 d,inAD4 a,AD1 b){d=a+AD4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opACpyD2(outAD2 d,inAD2 a){d=a;return d;} + AD3 opACpyD3(outAD3 d,inAD3 a){d=a;return d;} + AD4 opACpyD4(outAD4 d,inAD4 a){d=a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opALerpD2(outAD2 d,inAD2 a,inAD2 b,inAD2 c){d=ALerpD2(a,b,c);return d;} + AD3 opALerpD3(outAD3 d,inAD3 a,inAD3 b,inAD3 c){d=ALerpD3(a,b,c);return d;} + AD4 opALerpD4(outAD4 d,inAD4 a,inAD4 b,inAD4 c){d=ALerpD4(a,b,c);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opALerpOneD2(outAD2 d,inAD2 a,inAD2 b,AD1 c){d=ALerpD2(a,b,AD2_(c));return d;} + AD3 opALerpOneD3(outAD3 d,inAD3 a,inAD3 b,AD1 c){d=ALerpD3(a,b,AD3_(c));return d;} + AD4 opALerpOneD4(outAD4 d,inAD4 a,inAD4 b,AD1 c){d=ALerpD4(a,b,AD4_(c));return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMaxD2(outAD2 d,inAD2 a,inAD2 b){d=max(a,b);return d;} + AD3 opAMaxD3(outAD3 d,inAD3 a,inAD3 b){d=max(a,b);return d;} + AD4 opAMaxD4(outAD4 d,inAD4 a,inAD4 b){d=max(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMinD2(outAD2 d,inAD2 a,inAD2 b){d=min(a,b);return d;} + AD3 opAMinD3(outAD3 d,inAD3 a,inAD3 b){d=min(a,b);return d;} + AD4 opAMinD4(outAD4 d,inAD4 a,inAD4 b){d=min(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMulD2(outAD2 d,inAD2 a,inAD2 b){d=a*b;return d;} + AD3 opAMulD3(outAD3 d,inAD3 a,inAD3 b){d=a*b;return d;} + AD4 opAMulD4(outAD4 d,inAD4 a,inAD4 b){d=a*b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMulOneD2(outAD2 d,inAD2 a,AD1 b){d=a*AD2_(b);return d;} + AD3 opAMulOneD3(outAD3 d,inAD3 a,AD1 b){d=a*AD3_(b);return d;} + AD4 opAMulOneD4(outAD4 d,inAD4 a,AD1 b){d=a*AD4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opANegD2(outAD2 d,inAD2 a){d=-a;return d;} + AD3 opANegD3(outAD3 d,inAD3 a){d=-a;return d;} + AD4 opANegD4(outAD4 d,inAD4 a){d=-a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opARcpD2(outAD2 d,inAD2 a){d=ARcpD2(a);return d;} + AD3 opARcpD3(outAD3 d,inAD3 a){d=ARcpD3(a);return d;} + AD4 opARcpD4(outAD4 d,inAD4 a){d=ARcpD4(a);return d;} + #endif +//============================================================================================================================== + AF2 opAAbsF2(outAF2 d,inAF2 a){d=abs(a);return d;} + AF3 opAAbsF3(outAF3 d,inAF3 a){d=abs(a);return d;} + AF4 opAAbsF4(outAF4 d,inAF4 a){d=abs(a);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAAddF2(outAF2 d,inAF2 a,inAF2 b){d=a+b;return d;} + AF3 opAAddF3(outAF3 d,inAF3 a,inAF3 b){d=a+b;return d;} + AF4 opAAddF4(outAF4 d,inAF4 a,inAF4 b){d=a+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAAddOneF2(outAF2 d,inAF2 a,AF1 b){d=a+AF2_(b);return d;} + AF3 opAAddOneF3(outAF3 d,inAF3 a,AF1 b){d=a+AF3_(b);return d;} + AF4 opAAddOneF4(outAF4 d,inAF4 a,AF1 b){d=a+AF4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opACpyF2(outAF2 d,inAF2 a){d=a;return d;} + AF3 opACpyF3(outAF3 d,inAF3 a){d=a;return d;} + AF4 opACpyF4(outAF4 d,inAF4 a){d=a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opALerpF2(outAF2 d,inAF2 a,inAF2 b,inAF2 c){d=ALerpF2(a,b,c);return d;} + AF3 opALerpF3(outAF3 d,inAF3 a,inAF3 b,inAF3 c){d=ALerpF3(a,b,c);return d;} + AF4 opALerpF4(outAF4 d,inAF4 a,inAF4 b,inAF4 c){d=ALerpF4(a,b,c);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opALerpOneF2(outAF2 d,inAF2 a,inAF2 b,AF1 c){d=ALerpF2(a,b,AF2_(c));return d;} + AF3 opALerpOneF3(outAF3 d,inAF3 a,inAF3 b,AF1 c){d=ALerpF3(a,b,AF3_(c));return d;} + AF4 opALerpOneF4(outAF4 d,inAF4 a,inAF4 b,AF1 c){d=ALerpF4(a,b,AF4_(c));return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMaxF2(outAF2 d,inAF2 a,inAF2 b){d=max(a,b);return d;} + AF3 opAMaxF3(outAF3 d,inAF3 a,inAF3 b){d=max(a,b);return d;} + AF4 opAMaxF4(outAF4 d,inAF4 a,inAF4 b){d=max(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMinF2(outAF2 d,inAF2 a,inAF2 b){d=min(a,b);return d;} + AF3 opAMinF3(outAF3 d,inAF3 a,inAF3 b){d=min(a,b);return d;} + AF4 opAMinF4(outAF4 d,inAF4 a,inAF4 b){d=min(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMulF2(outAF2 d,inAF2 a,inAF2 b){d=a*b;return d;} + AF3 opAMulF3(outAF3 d,inAF3 a,inAF3 b){d=a*b;return d;} + AF4 opAMulF4(outAF4 d,inAF4 a,inAF4 b){d=a*b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMulOneF2(outAF2 d,inAF2 a,AF1 b){d=a*AF2_(b);return d;} + AF3 opAMulOneF3(outAF3 d,inAF3 a,AF1 b){d=a*AF3_(b);return d;} + AF4 opAMulOneF4(outAF4 d,inAF4 a,AF1 b){d=a*AF4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opANegF2(outAF2 d,inAF2 a){d=-a;return d;} + AF3 opANegF3(outAF3 d,inAF3 a){d=-a;return d;} + AF4 opANegF4(outAF4 d,inAF4 a){d=-a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opARcpF2(outAF2 d,inAF2 a){d=ARcpF2(a);return d;} + AF3 opARcpF3(outAF3 d,inAF3 a){d=ARcpF3(a);return d;} + AF4 opARcpF4(outAF4 d,inAF4 a){d=ARcpF4(a);return d;} +#endif + +#define FSR_EASU_F 1 +AU4 con0, con1, con2, con3; +float srcW, srcH, dstW, dstH; +vec2 bLeft, tRight; + +AF2 translate(AF2 pos) { + return AF2(pos.x * scaleX, pos.y * scaleY); +} + +void setBounds(vec2 bottomLeft, vec2 topRight) { + bLeft = bottomLeft; + tRight = topRight; +} + +AF4 FsrEasuRF(AF2 p) { AF4 res = textureGather(Source, translate(p), 0); return res; } +AF4 FsrEasuGF(AF2 p) { AF4 res = textureGather(Source, translate(p), 1); return res; } +AF4 FsrEasuBF(AF2 p) { AF4 res = textureGather(Source, translate(p), 2); return res; } + +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// AMD FidelityFX SUPER RESOLUTION [FSR 1] ::: SPATIAL SCALING & EXTRAS - v1.20210629 +// +// +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// FidelityFX Super Resolution Sample +// +// Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved. +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files(the "Software"), to deal +// in the Software without restriction, including without limitation the rights +// to use, copy, modify, merge, publish, distribute, sublicense, and / or sell +// copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions : +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN +// THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// ABOUT +// ===== +// FSR is a collection of algorithms relating to generating a higher resolution image. +// This specific header focuses on single-image non-temporal image scaling, and related tools. +// +// The core functions are EASU and RCAS: +// [EASU] Edge Adaptive Spatial Upsampling ....... 1x to 4x area range spatial scaling, clamped adaptive elliptical filter. +// [RCAS] Robust Contrast Adaptive Sharpening .... A non-scaling variation on CAS. +// RCAS needs to be applied after EASU as a separate pass. +// +// Optional utility functions are: +// [LFGA] Linear Film Grain Applicator ........... Tool to apply film grain after scaling. +// [SRTM] Simple Reversible Tone-Mapper .......... Linear HDR {0 to FP16_MAX} to {0 to 1} and back. +// [TEPD] Temporal Energy Preserving Dither ...... Temporally energy preserving dithered {0 to 1} linear to gamma 2.0 conversion. +// See each individual sub-section for inline documentation. +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// FUNCTION PERMUTATIONS +// ===================== +// *F() ..... Single item computation with 32-bit. +// *H() ..... Single item computation with 16-bit, with packing (aka two 16-bit ops in parallel) when possible. +// *Hx2() ... Processing two items in parallel with 16-bit, easier packing. +// Not all interfaces in this file have a *Hx2() form. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [EASU] EDGE ADAPTIVE SPATIAL UPSAMPLING +// +//------------------------------------------------------------------------------------------------------------------------------ +// EASU provides a high quality spatial-only scaling at relatively low cost. +// Meaning EASU is appropiate for laptops and other low-end GPUs. +// Quality from 1x to 4x area scaling is good. +//------------------------------------------------------------------------------------------------------------------------------ +// The scalar uses a modified fast approximation to the standard lanczos(size=2) kernel. +// EASU runs in a single pass, so it applies a directionally and anisotropically adaptive radial lanczos. +// This is also kept as simple as possible to have minimum runtime. +//------------------------------------------------------------------------------------------------------------------------------ +// The lanzcos filter has negative lobes, so by itself it will introduce ringing. +// To remove all ringing, the algorithm uses the nearest 2x2 input texels as a neighborhood, +// and limits output to the minimum and maximum of that neighborhood. +//------------------------------------------------------------------------------------------------------------------------------ +// Input image requirements: +// +// Color needs to be encoded as 3 channel[red, green, blue](e.g.XYZ not supported) +// Each channel needs to be in the range[0, 1] +// Any color primaries are supported +// Display / tonemapping curve needs to be as if presenting to sRGB display or similar(e.g.Gamma 2.0) +// There should be no banding in the input +// There should be no high amplitude noise in the input +// There should be no noise in the input that is not at input pixel granularity +// For performance purposes, use 32bpp formats +//------------------------------------------------------------------------------------------------------------------------------ +// Best to apply EASU at the end of the frame after tonemapping +// but before film grain or composite of the UI. +//------------------------------------------------------------------------------------------------------------------------------ +// Example of including this header for D3D HLSL : +// +// #define A_GPU 1 +// #define A_HLSL 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of including this header for Vulkan GLSL : +// +// #define A_GPU 1 +// #define A_GLSL 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of including this header for Vulkan HLSL : +// +// #define A_GPU 1 +// #define A_HLSL 1 +// #define A_HLSL_6_2 1 +// #define A_NO_16_BIT_CAST 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of declaring the required input callbacks for GLSL : +// The callbacks need to gather4 for each color channel using the specified texture coordinate 'p'. +// EASU uses gather4 to reduce position computation logic and for free Arrays of Structures to Structures of Arrays conversion. +// +// AH4 FsrEasuRH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,0));} +// AH4 FsrEasuGH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,1));} +// AH4 FsrEasuBH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,2));} +// ... +// The FsrEasuCon function needs to be called from the CPU or GPU to set up constants. +// The difference in viewport and input image size is there to support Dynamic Resolution Scaling. +// To use FsrEasuCon() on the CPU, define A_CPU before including ffx_a and ffx_fsr1. +// Including a GPU example here, the 'con0' through 'con3' values would be stored out to a constant buffer. +// AU4 con0,con1,con2,con3; +// FsrEasuCon(con0,con1,con2,con3, +// 1920.0,1080.0, // Viewport size (top left aligned) in the input image which is to be scaled. +// 3840.0,2160.0, // The size of the input image. +// 2560.0,1440.0); // The output resolution. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CONSTANT SETUP +//============================================================================================================================== +// Call to setup required constant values (works on CPU or GPU). +A_STATIC void FsrEasuCon( +outAU4 con0, +outAU4 con1, +outAU4 con2, +outAU4 con3, +// This the rendered image resolution being upscaled +AF1 inputViewportInPixelsX, +AF1 inputViewportInPixelsY, +// This is the resolution of the resource containing the input image (useful for dynamic resolution) +AF1 inputSizeInPixelsX, +AF1 inputSizeInPixelsY, +// This is the display resolution which the input image gets upscaled to +AF1 outputSizeInPixelsX, +AF1 outputSizeInPixelsY){ + // Output integer position to a pixel position in viewport. + con0[0]=AU1_AF1(inputViewportInPixelsX*ARcpF1(outputSizeInPixelsX)); + con0[1]=AU1_AF1(inputViewportInPixelsY*ARcpF1(outputSizeInPixelsY)); + con0[2]=AU1_AF1(AF1_(0.5)*inputViewportInPixelsX*ARcpF1(outputSizeInPixelsX)-AF1_(0.5)); + con0[3]=AU1_AF1(AF1_(0.5)*inputViewportInPixelsY*ARcpF1(outputSizeInPixelsY)-AF1_(0.5)); + // Viewport pixel position to normalized image space. + // This is used to get upper-left of 'F' tap. + con1[0]=AU1_AF1(ARcpF1(inputSizeInPixelsX)); + con1[1]=AU1_AF1(ARcpF1(inputSizeInPixelsY)); + // Centers of gather4, first offset from upper-left of 'F'. + // +---+---+ + // | | | + // +--(0)--+ + // | b | c | + // +---F---+---+---+ + // | e | f | g | h | + // +--(1)--+--(2)--+ + // | i | j | k | l | + // +---+---+---+---+ + // | n | o | + // +--(3)--+ + // | | | + // +---+---+ + con1[2]=AU1_AF1(AF1_( 1.0)*ARcpF1(inputSizeInPixelsX)); + con1[3]=AU1_AF1(AF1_(-1.0)*ARcpF1(inputSizeInPixelsY)); + // These are from (0) instead of 'F'. + con2[0]=AU1_AF1(AF1_(-1.0)*ARcpF1(inputSizeInPixelsX)); + con2[1]=AU1_AF1(AF1_( 2.0)*ARcpF1(inputSizeInPixelsY)); + con2[2]=AU1_AF1(AF1_( 1.0)*ARcpF1(inputSizeInPixelsX)); + con2[3]=AU1_AF1(AF1_( 2.0)*ARcpF1(inputSizeInPixelsY)); + con3[0]=AU1_AF1(AF1_( 0.0)*ARcpF1(inputSizeInPixelsX)); + con3[1]=AU1_AF1(AF1_( 4.0)*ARcpF1(inputSizeInPixelsY)); + con3[2]=con3[3]=0;} + +//If the an offset into the input image resource +A_STATIC void FsrEasuConOffset( + outAU4 con0, + outAU4 con1, + outAU4 con2, + outAU4 con3, + // This the rendered image resolution being upscaled + AF1 inputViewportInPixelsX, + AF1 inputViewportInPixelsY, + // This is the resolution of the resource containing the input image (useful for dynamic resolution) + AF1 inputSizeInPixelsX, + AF1 inputSizeInPixelsY, + // This is the display resolution which the input image gets upscaled to + AF1 outputSizeInPixelsX, + AF1 outputSizeInPixelsY, + // This is the input image offset into the resource containing it (useful for dynamic resolution) + AF1 inputOffsetInPixelsX, + AF1 inputOffsetInPixelsY) { + FsrEasuCon(con0, con1, con2, con3, inputViewportInPixelsX, inputViewportInPixelsY, inputSizeInPixelsX, inputSizeInPixelsY, outputSizeInPixelsX, outputSizeInPixelsY); + con0[2] = AU1_AF1(AF1_(0.5) * inputViewportInPixelsX * ARcpF1(outputSizeInPixelsX) - AF1_(0.5) + inputOffsetInPixelsX); + con0[3] = AU1_AF1(AF1_(0.5) * inputViewportInPixelsY * ARcpF1(outputSizeInPixelsY) - AF1_(0.5) + inputOffsetInPixelsY); +} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 32-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(FSR_EASU_F) + // Input callback prototypes, need to be implemented by calling shader + AF4 FsrEasuRF(AF2 p); + AF4 FsrEasuGF(AF2 p); + AF4 FsrEasuBF(AF2 p); +//------------------------------------------------------------------------------------------------------------------------------ + // Filtering for a given tap for the scalar. + void FsrEasuTapF( + inout AF3 aC, // Accumulated color, with negative lobe. + inout AF1 aW, // Accumulated weight. + AF2 off, // Pixel offset from resolve position to tap. + AF2 dir, // Gradient direction. + AF2 len, // Length. + AF1 lob, // Negative lobe strength. + AF1 clp, // Clipping point. + AF3 c){ // Tap color. + // Rotate offset by direction. + AF2 v; + v.x=(off.x*( dir.x))+(off.y*dir.y); + v.y=(off.x*(-dir.y))+(off.y*dir.x); + // Anisotropy. + v*=len; + // Compute distance^2. + AF1 d2=v.x*v.x+v.y*v.y; + // Limit to the window as at corner, 2 taps can easily be outside. + d2=min(d2,clp); + // Approximation of lancos2 without sin() or rcp(), or sqrt() to get x. + // (25/16 * (2/5 * x^2 - 1)^2 - (25/16 - 1)) * (1/4 * x^2 - 1)^2 + // |_______________________________________| |_______________| + // base window + // The general form of the 'base' is, + // (a*(b*x^2-1)^2-(a-1)) + // Where 'a=1/(2*b-b^2)' and 'b' moves around the negative lobe. + AF1 wB=AF1_(2.0/5.0)*d2+AF1_(-1.0); + AF1 wA=lob*d2+AF1_(-1.0); + wB*=wB; + wA*=wA; + wB=AF1_(25.0/16.0)*wB+AF1_(-(25.0/16.0-1.0)); + AF1 w=wB*wA; + // Do weighted average. + aC+=c*w;aW+=w;} +//------------------------------------------------------------------------------------------------------------------------------ + // Accumulate direction and length. + void FsrEasuSetF( + inout AF2 dir, + inout AF1 len, + AF2 pp, + AP1 biS,AP1 biT,AP1 biU,AP1 biV, + AF1 lA,AF1 lB,AF1 lC,AF1 lD,AF1 lE){ + // Compute bilinear weight, branches factor out as predicates are compiler time immediates. + // s t + // u v + AF1 w = AF1_(0.0); + if(biS)w=(AF1_(1.0)-pp.x)*(AF1_(1.0)-pp.y); + if(biT)w= pp.x *(AF1_(1.0)-pp.y); + if(biU)w=(AF1_(1.0)-pp.x)* pp.y ; + if(biV)w= pp.x * pp.y ; + // Direction is the '+' diff. + // a + // b c d + // e + // Then takes magnitude from abs average of both sides of 'c'. + // Length converts gradient reversal to 0, smoothly to non-reversal at 1, shaped, then adding horz and vert terms. + AF1 dc=lD-lC; + AF1 cb=lC-lB; + AF1 lenX=max(abs(dc),abs(cb)); + lenX=APrxLoRcpF1(lenX); + AF1 dirX=lD-lB; + dir.x+=dirX*w; + lenX=ASatF1(abs(dirX)*lenX); + lenX*=lenX; + len+=lenX*w; + // Repeat for the y axis. + AF1 ec=lE-lC; + AF1 ca=lC-lA; + AF1 lenY=max(abs(ec),abs(ca)); + lenY=APrxLoRcpF1(lenY); + AF1 dirY=lE-lA; + dir.y+=dirY*w; + lenY=ASatF1(abs(dirY)*lenY); + lenY*=lenY; + len+=lenY*w;} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrEasuF( + out AF3 pix, + AU2 ip, // Integer pixel position in output. + AU4 con0, // Constants generated by FsrEasuCon(). + AU4 con1, + AU4 con2, + AU4 con3){ +//------------------------------------------------------------------------------------------------------------------------------ + // Get position of 'f'. + AF2 pp=AF2(ip)*AF2_AU2(con0.xy)+AF2_AU2(con0.zw); + AF2 fp=floor(pp); + pp-=fp; +//------------------------------------------------------------------------------------------------------------------------------ + // 12-tap kernel. + // b c + // e f g h + // i j k l + // n o + // Gather 4 ordering. + // a b + // r g + // For packed FP16, need either {rg} or {ab} so using the following setup for gather in all versions, + // a b <- unused (z) + // r g + // a b a b + // r g r g + // a b + // r g <- unused (z) + // Allowing dead-code removal to remove the 'z's. + AF2 p0=fp*AF2_AU2(con1.xy)+AF2_AU2(con1.zw); + // These are from p0 to avoid pulling two constants on pre-Navi hardware. + AF2 p1=p0+AF2_AU2(con2.xy); + AF2 p2=p0+AF2_AU2(con2.zw); + AF2 p3=p0+AF2_AU2(con3.xy); + AF4 bczzR=FsrEasuRF(p0); + AF4 bczzG=FsrEasuGF(p0); + AF4 bczzB=FsrEasuBF(p0); + AF4 ijfeR=FsrEasuRF(p1); + AF4 ijfeG=FsrEasuGF(p1); + AF4 ijfeB=FsrEasuBF(p1); + AF4 klhgR=FsrEasuRF(p2); + AF4 klhgG=FsrEasuGF(p2); + AF4 klhgB=FsrEasuBF(p2); + AF4 zzonR=FsrEasuRF(p3); + AF4 zzonG=FsrEasuGF(p3); + AF4 zzonB=FsrEasuBF(p3); +//------------------------------------------------------------------------------------------------------------------------------ + // Simplest multi-channel approximate luma possible (luma times 2, in 2 FMA/MAD). + AF4 bczzL=bczzB*AF4_(0.5)+(bczzR*AF4_(0.5)+bczzG); + AF4 ijfeL=ijfeB*AF4_(0.5)+(ijfeR*AF4_(0.5)+ijfeG); + AF4 klhgL=klhgB*AF4_(0.5)+(klhgR*AF4_(0.5)+klhgG); + AF4 zzonL=zzonB*AF4_(0.5)+(zzonR*AF4_(0.5)+zzonG); + // Rename. + AF1 bL=bczzL.x; + AF1 cL=bczzL.y; + AF1 iL=ijfeL.x; + AF1 jL=ijfeL.y; + AF1 fL=ijfeL.z; + AF1 eL=ijfeL.w; + AF1 kL=klhgL.x; + AF1 lL=klhgL.y; + AF1 hL=klhgL.z; + AF1 gL=klhgL.w; + AF1 oL=zzonL.z; + AF1 nL=zzonL.w; + // Accumulate for bilinear interpolation. + AF2 dir=AF2_(0.0); + AF1 len=AF1_(0.0); + FsrEasuSetF(dir,len,pp,true, false,false,false,bL,eL,fL,gL,jL); + FsrEasuSetF(dir,len,pp,false,true ,false,false,cL,fL,gL,hL,kL); + FsrEasuSetF(dir,len,pp,false,false,true ,false,fL,iL,jL,kL,nL); + FsrEasuSetF(dir,len,pp,false,false,false,true ,gL,jL,kL,lL,oL); +//------------------------------------------------------------------------------------------------------------------------------ + // Normalize with approximation, and cleanup close to zero. + AF2 dir2=dir*dir; + AF1 dirR=dir2.x+dir2.y; + AP1 zro=dirR w = -m/(n+e+w+s) +// 1 == (w*(n+e+w+s)+m)/(4*w+1) -> w = (1-m)/(n+e+w+s-4*1) +// Then chooses the 'w' which results in no clipping, limits 'w', and multiplies by the 'sharp' amount. +// This solution above has issues with MSAA input as the steps along the gradient cause edge detection issues. +// So RCAS uses 4x the maximum and 4x the minimum (depending on equation)in place of the individual taps. +// As well as switching from 'm' to either the minimum or maximum (depending on side), to help in energy conservation. +// This stabilizes RCAS. +// RCAS does a simple highpass which is normalized against the local contrast then shaped, +// 0.25 +// 0.25 -1 0.25 +// 0.25 +// This is used as a noise detection filter, to reduce the effect of RCAS on grain, and focus on real edges. +// +// GLSL example for the required callbacks : +// +// AH4 FsrRcasLoadH(ASW2 p){return AH4(imageLoad(imgSrc,ASU2(p)));} +// void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b) +// { +// //do any simple input color conversions here or leave empty if none needed +// } +// +// FsrRcasCon need to be called from the CPU or GPU to set up constants. +// Including a GPU example here, the 'con' value would be stored out to a constant buffer. +// +// AU4 con; +// FsrRcasCon(con, +// 0.0); // The scale is {0.0 := maximum sharpness, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. +// --------------- +// RCAS sharpening supports a CAS-like pass-through alpha via, +// #define FSR_RCAS_PASSTHROUGH_ALPHA 1 +// RCAS also supports a define to enable a more expensive path to avoid some sharpening of noise. +// Would suggest it is better to apply film grain after RCAS sharpening (and after scaling) instead of using this define, +// #define FSR_RCAS_DENOISE 1 +//============================================================================================================================== +// This is set at the limit of providing unnatural results for sharpening. +#define FSR_RCAS_LIMIT (0.25-(1.0/16.0)) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CONSTANT SETUP +//============================================================================================================================== +// Call to setup required constant values (works on CPU or GPU). +A_STATIC void FsrRcasCon( +outAU4 con, +// The scale is {0.0 := maximum, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. +AF1 sharpness){ + // Transform from stops to linear value. + sharpness=AExp2F1(-sharpness); + varAF2(hSharp)=initAF2(sharpness,sharpness); + con[0]=AU1_AF1(sharpness); + con[1]=AU1_AH2_AF2(hSharp); + con[2]=0; + con[3]=0;} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 32-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(FSR_RCAS_F) + // Input callback prototypes that need to be implemented by calling shader + AF4 FsrRcasLoadF(ASU2 p); + void FsrRcasInputF(inout AF1 r,inout AF1 g,inout AF1 b); +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasF( + out AF1 pixR, // Output values, non-vector so port between RcasFilter() and RcasFilterH() is easy. + out AF1 pixG, + out AF1 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AF1 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // Algorithm uses minimal 3x3 pixel neighborhood. + // b + // d e f + // h + ASU2 sp=ASU2(ip); + AF3 b=FsrRcasLoadF(sp+ASU2( 0,-1)).rgb; + AF3 d=FsrRcasLoadF(sp+ASU2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AF4 ee=FsrRcasLoadF(sp); + AF3 e=ee.rgb;pixA=ee.a; + #else + AF3 e=FsrRcasLoadF(sp).rgb; + #endif + AF3 f=FsrRcasLoadF(sp+ASU2( 1, 0)).rgb; + AF3 h=FsrRcasLoadF(sp+ASU2( 0, 1)).rgb; + // Rename (32-bit) or regroup (16-bit). + AF1 bR=b.r; + AF1 bG=b.g; + AF1 bB=b.b; + AF1 dR=d.r; + AF1 dG=d.g; + AF1 dB=d.b; + AF1 eR=e.r; + AF1 eG=e.g; + AF1 eB=e.b; + AF1 fR=f.r; + AF1 fG=f.g; + AF1 fB=f.b; + AF1 hR=h.r; + AF1 hG=h.g; + AF1 hB=h.b; + // Run optional input transform. + FsrRcasInputF(bR,bG,bB); + FsrRcasInputF(dR,dG,dB); + FsrRcasInputF(eR,eG,eB); + FsrRcasInputF(fR,fG,fB); + FsrRcasInputF(hR,hG,hB); + // Luma times 2. + AF1 bL=bB*AF1_(0.5)+(bR*AF1_(0.5)+bG); + AF1 dL=dB*AF1_(0.5)+(dR*AF1_(0.5)+dG); + AF1 eL=eB*AF1_(0.5)+(eR*AF1_(0.5)+eG); + AF1 fL=fB*AF1_(0.5)+(fR*AF1_(0.5)+fG); + AF1 hL=hB*AF1_(0.5)+(hR*AF1_(0.5)+hG); + // Noise detection. + AF1 nz=AF1_(0.25)*bL+AF1_(0.25)*dL+AF1_(0.25)*fL+AF1_(0.25)*hL-eL; + nz=ASatF1(abs(nz)*APrxMedRcpF1(AMax3F1(AMax3F1(bL,dL,eL),fL,hL)-AMin3F1(AMin3F1(bL,dL,eL),fL,hL))); + nz=AF1_(-0.5)*nz+AF1_(1.0); + // Min and max of ring. + AF1 mn4R=min(AMin3F1(bR,dR,fR),hR); + AF1 mn4G=min(AMin3F1(bG,dG,fG),hG); + AF1 mn4B=min(AMin3F1(bB,dB,fB),hB); + AF1 mx4R=max(AMax3F1(bR,dR,fR),hR); + AF1 mx4G=max(AMax3F1(bG,dG,fG),hG); + AF1 mx4B=max(AMax3F1(bB,dB,fB),hB); + // Immediate constants for peak range. + AF2 peakC=AF2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AF1 hitMinR=min(mn4R,eR)*ARcpF1(AF1_(4.0)*mx4R); + AF1 hitMinG=min(mn4G,eG)*ARcpF1(AF1_(4.0)*mx4G); + AF1 hitMinB=min(mn4B,eB)*ARcpF1(AF1_(4.0)*mx4B); + AF1 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpF1(AF1_(4.0)*mn4R+peakC.y); + AF1 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpF1(AF1_(4.0)*mn4G+peakC.y); + AF1 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpF1(AF1_(4.0)*mn4B+peakC.y); + AF1 lobeR=max(-hitMinR,hitMaxR); + AF1 lobeG=max(-hitMinG,hitMaxG); + AF1 lobeB=max(-hitMinB,hitMaxB); + AF1 lobe=max(AF1_(-FSR_RCAS_LIMIT),min(AMax3F1(lobeR,lobeG,lobeB),AF1_(0.0)))*AF1_AU1(con.x); + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AF1 rcpL=APrxMedRcpF1(AF1_(4.0)*lobe+AF1_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL; + return;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 16-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_RCAS_H) + // Input callback prototypes that need to be implemented by calling shader + AH4 FsrRcasLoadH(ASW2 p); + void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b); +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasH( + out AH1 pixR, // Output values, non-vector so port between RcasFilter() and RcasFilterH() is easy. + out AH1 pixG, + out AH1 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AH1 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // Sharpening algorithm uses minimal 3x3 pixel neighborhood. + // b + // d e f + // h + ASW2 sp=ASW2(ip); + AH3 b=FsrRcasLoadH(sp+ASW2( 0,-1)).rgb; + AH3 d=FsrRcasLoadH(sp+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee=FsrRcasLoadH(sp); + AH3 e=ee.rgb;pixA=ee.a; + #else + AH3 e=FsrRcasLoadH(sp).rgb; + #endif + AH3 f=FsrRcasLoadH(sp+ASW2( 1, 0)).rgb; + AH3 h=FsrRcasLoadH(sp+ASW2( 0, 1)).rgb; + // Rename (32-bit) or regroup (16-bit). + AH1 bR=b.r; + AH1 bG=b.g; + AH1 bB=b.b; + AH1 dR=d.r; + AH1 dG=d.g; + AH1 dB=d.b; + AH1 eR=e.r; + AH1 eG=e.g; + AH1 eB=e.b; + AH1 fR=f.r; + AH1 fG=f.g; + AH1 fB=f.b; + AH1 hR=h.r; + AH1 hG=h.g; + AH1 hB=h.b; + // Run optional input transform. + FsrRcasInputH(bR,bG,bB); + FsrRcasInputH(dR,dG,dB); + FsrRcasInputH(eR,eG,eB); + FsrRcasInputH(fR,fG,fB); + FsrRcasInputH(hR,hG,hB); + // Luma times 2. + AH1 bL=bB*AH1_(0.5)+(bR*AH1_(0.5)+bG); + AH1 dL=dB*AH1_(0.5)+(dR*AH1_(0.5)+dG); + AH1 eL=eB*AH1_(0.5)+(eR*AH1_(0.5)+eG); + AH1 fL=fB*AH1_(0.5)+(fR*AH1_(0.5)+fG); + AH1 hL=hB*AH1_(0.5)+(hR*AH1_(0.5)+hG); + // Noise detection. + AH1 nz=AH1_(0.25)*bL+AH1_(0.25)*dL+AH1_(0.25)*fL+AH1_(0.25)*hL-eL; + nz=ASatH1(abs(nz)*APrxMedRcpH1(AMax3H1(AMax3H1(bL,dL,eL),fL,hL)-AMin3H1(AMin3H1(bL,dL,eL),fL,hL))); + nz=AH1_(-0.5)*nz+AH1_(1.0); + // Min and max of ring. + AH1 mn4R=min(AMin3H1(bR,dR,fR),hR); + AH1 mn4G=min(AMin3H1(bG,dG,fG),hG); + AH1 mn4B=min(AMin3H1(bB,dB,fB),hB); + AH1 mx4R=max(AMax3H1(bR,dR,fR),hR); + AH1 mx4G=max(AMax3H1(bG,dG,fG),hG); + AH1 mx4B=max(AMax3H1(bB,dB,fB),hB); + // Immediate constants for peak range. + AH2 peakC=AH2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AH1 hitMinR=min(mn4R,eR)*ARcpH1(AH1_(4.0)*mx4R); + AH1 hitMinG=min(mn4G,eG)*ARcpH1(AH1_(4.0)*mx4G); + AH1 hitMinB=min(mn4B,eB)*ARcpH1(AH1_(4.0)*mx4B); + AH1 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpH1(AH1_(4.0)*mn4R+peakC.y); + AH1 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpH1(AH1_(4.0)*mn4G+peakC.y); + AH1 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpH1(AH1_(4.0)*mn4B+peakC.y); + AH1 lobeR=max(-hitMinR,hitMaxR); + AH1 lobeG=max(-hitMinG,hitMaxG); + AH1 lobeB=max(-hitMinB,hitMaxB); + AH1 lobe=max(AH1_(-FSR_RCAS_LIMIT),min(AMax3H1(lobeR,lobeG,lobeB),AH1_(0.0)))*AH2_AU1(con.y).x; + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AH1 rcpL=APrxMedRcpH1(AH1_(4.0)*lobe+AH1_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PACKED 16-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_RCAS_HX2) + // Input callback prototypes that need to be implemented by the calling shader + AH4 FsrRcasLoadHx2(ASW2 p); + void FsrRcasInputHx2(inout AH2 r,inout AH2 g,inout AH2 b); +//------------------------------------------------------------------------------------------------------------------------------ + // Can be used to convert from packed Structures of Arrays to Arrays of Structures for store. + void FsrRcasDepackHx2(out AH4 pix0,out AH4 pix1,AH2 pixR,AH2 pixG,AH2 pixB){ + #ifdef A_HLSL + // Invoke a slower path for DX only, since it won't allow uninitialized values. + pix0.a=pix1.a=0.0; + #endif + pix0.rgb=AH3(pixR.x,pixG.x,pixB.x); + pix1.rgb=AH3(pixR.y,pixG.y,pixB.y);} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasHx2( + // Output values are for 2 8x8 tiles in a 16x8 region. + // pix.x = left 8x8 tile + // pix.y = right 8x8 tile + // This enables later processing to easily be packed as well. + out AH2 pixR, + out AH2 pixG, + out AH2 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AH2 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // No scaling algorithm uses minimal 3x3 pixel neighborhood. + ASW2 sp0=ASW2(ip); + AH3 b0=FsrRcasLoadHx2(sp0+ASW2( 0,-1)).rgb; + AH3 d0=FsrRcasLoadHx2(sp0+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee0=FsrRcasLoadHx2(sp0); + AH3 e0=ee0.rgb;pixA.r=ee0.a; + #else + AH3 e0=FsrRcasLoadHx2(sp0).rgb; + #endif + AH3 f0=FsrRcasLoadHx2(sp0+ASW2( 1, 0)).rgb; + AH3 h0=FsrRcasLoadHx2(sp0+ASW2( 0, 1)).rgb; + ASW2 sp1=sp0+ASW2(8,0); + AH3 b1=FsrRcasLoadHx2(sp1+ASW2( 0,-1)).rgb; + AH3 d1=FsrRcasLoadHx2(sp1+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee1=FsrRcasLoadHx2(sp1); + AH3 e1=ee1.rgb;pixA.g=ee1.a; + #else + AH3 e1=FsrRcasLoadHx2(sp1).rgb; + #endif + AH3 f1=FsrRcasLoadHx2(sp1+ASW2( 1, 0)).rgb; + AH3 h1=FsrRcasLoadHx2(sp1+ASW2( 0, 1)).rgb; + // Arrays of Structures to Structures of Arrays conversion. + AH2 bR=AH2(b0.r,b1.r); + AH2 bG=AH2(b0.g,b1.g); + AH2 bB=AH2(b0.b,b1.b); + AH2 dR=AH2(d0.r,d1.r); + AH2 dG=AH2(d0.g,d1.g); + AH2 dB=AH2(d0.b,d1.b); + AH2 eR=AH2(e0.r,e1.r); + AH2 eG=AH2(e0.g,e1.g); + AH2 eB=AH2(e0.b,e1.b); + AH2 fR=AH2(f0.r,f1.r); + AH2 fG=AH2(f0.g,f1.g); + AH2 fB=AH2(f0.b,f1.b); + AH2 hR=AH2(h0.r,h1.r); + AH2 hG=AH2(h0.g,h1.g); + AH2 hB=AH2(h0.b,h1.b); + // Run optional input transform. + FsrRcasInputHx2(bR,bG,bB); + FsrRcasInputHx2(dR,dG,dB); + FsrRcasInputHx2(eR,eG,eB); + FsrRcasInputHx2(fR,fG,fB); + FsrRcasInputHx2(hR,hG,hB); + // Luma times 2. + AH2 bL=bB*AH2_(0.5)+(bR*AH2_(0.5)+bG); + AH2 dL=dB*AH2_(0.5)+(dR*AH2_(0.5)+dG); + AH2 eL=eB*AH2_(0.5)+(eR*AH2_(0.5)+eG); + AH2 fL=fB*AH2_(0.5)+(fR*AH2_(0.5)+fG); + AH2 hL=hB*AH2_(0.5)+(hR*AH2_(0.5)+hG); + // Noise detection. + AH2 nz=AH2_(0.25)*bL+AH2_(0.25)*dL+AH2_(0.25)*fL+AH2_(0.25)*hL-eL; + nz=ASatH2(abs(nz)*APrxMedRcpH2(AMax3H2(AMax3H2(bL,dL,eL),fL,hL)-AMin3H2(AMin3H2(bL,dL,eL),fL,hL))); + nz=AH2_(-0.5)*nz+AH2_(1.0); + // Min and max of ring. + AH2 mn4R=min(AMin3H2(bR,dR,fR),hR); + AH2 mn4G=min(AMin3H2(bG,dG,fG),hG); + AH2 mn4B=min(AMin3H2(bB,dB,fB),hB); + AH2 mx4R=max(AMax3H2(bR,dR,fR),hR); + AH2 mx4G=max(AMax3H2(bG,dG,fG),hG); + AH2 mx4B=max(AMax3H2(bB,dB,fB),hB); + // Immediate constants for peak range. + AH2 peakC=AH2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AH2 hitMinR=min(mn4R,eR)*ARcpH2(AH2_(4.0)*mx4R); + AH2 hitMinG=min(mn4G,eG)*ARcpH2(AH2_(4.0)*mx4G); + AH2 hitMinB=min(mn4B,eB)*ARcpH2(AH2_(4.0)*mx4B); + AH2 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpH2(AH2_(4.0)*mn4R+peakC.y); + AH2 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpH2(AH2_(4.0)*mn4G+peakC.y); + AH2 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpH2(AH2_(4.0)*mn4B+peakC.y); + AH2 lobeR=max(-hitMinR,hitMaxR); + AH2 lobeG=max(-hitMinG,hitMaxG); + AH2 lobeB=max(-hitMinB,hitMaxB); + AH2 lobe=max(AH2_(-FSR_RCAS_LIMIT),min(AMax3H2(lobeR,lobeG,lobeB),AH2_(0.0)))*AH2_(AH2_AU1(con.y).x); + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AH2 rcpL=APrxMedRcpH2(AH2_(4.0)*lobe+AH2_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [LFGA] LINEAR FILM GRAIN APPLICATOR +// +//------------------------------------------------------------------------------------------------------------------------------ +// Adding output-resolution film grain after scaling is a good way to mask both rendering and scaling artifacts. +// Suggest using tiled blue noise as film grain input, with peak noise frequency set for a specific look and feel. +// The 'Lfga*()' functions provide a convenient way to introduce grain. +// These functions limit grain based on distance to signal limits. +// This is done so that the grain is temporally energy preserving, and thus won't modify image tonality. +// Grain application should be done in a linear colorspace. +// The grain should be temporally changing, but have a temporal sum per pixel that adds to zero (non-biased). +//------------------------------------------------------------------------------------------------------------------------------ +// Usage, +// FsrLfga*( +// color, // In/out linear colorspace color {0 to 1} ranged. +// grain, // Per pixel grain texture value {-0.5 to 0.5} ranged, input is 3-channel to support colored grain. +// amount); // Amount of grain (0 to 1} ranged. +//------------------------------------------------------------------------------------------------------------------------------ +// Example if grain texture is monochrome: 'FsrLfgaF(color,AF3_(grain),amount)' +//============================================================================================================================== +#if defined(A_GPU) + // Maximum grain is the minimum distance to the signal limit. + void FsrLfgaF(inout AF3 c,AF3 t,AF1 a){c+=(t*AF3_(a))*min(AF3_(1.0)-c,c);} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + // Half precision version (slower). + void FsrLfgaH(inout AH3 c,AH3 t,AH1 a){c+=(t*AH3_(a))*min(AH3_(1.0)-c,c);} +//------------------------------------------------------------------------------------------------------------------------------ + // Packed half precision version (faster). + void FsrLfgaHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 tR,AH2 tG,AH2 tB,AH1 a){ + cR+=(tR*AH2_(a))*min(AH2_(1.0)-cR,cR);cG+=(tG*AH2_(a))*min(AH2_(1.0)-cG,cG);cB+=(tB*AH2_(a))*min(AH2_(1.0)-cB,cB);} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [SRTM] SIMPLE REVERSIBLE TONE-MAPPER +// +//------------------------------------------------------------------------------------------------------------------------------ +// This provides a way to take linear HDR color {0 to FP16_MAX} and convert it into a temporary {0 to 1} ranged post-tonemapped linear. +// The tonemapper preserves RGB ratio, which helps maintain HDR color bleed during filtering. +//------------------------------------------------------------------------------------------------------------------------------ +// Reversible tonemapper usage, +// FsrSrtm*(color); // {0 to FP16_MAX} converted to {0 to 1}. +// FsrSrtmInv*(color); // {0 to 1} converted into {0 to 32768, output peak safe for FP16}. +//============================================================================================================================== +#if defined(A_GPU) + void FsrSrtmF(inout AF3 c){c*=AF3_(ARcpF1(AMax3F1(c.r,c.g,c.b)+AF1_(1.0)));} + // The extra max solves the c=1.0 case (which is a /0). + void FsrSrtmInvF(inout AF3 c){c*=AF3_(ARcpF1(max(AF1_(1.0/32768.0),AF1_(1.0)-AMax3F1(c.r,c.g,c.b))));} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + void FsrSrtmH(inout AH3 c){c*=AH3_(ARcpH1(AMax3H1(c.r,c.g,c.b)+AH1_(1.0)));} + void FsrSrtmInvH(inout AH3 c){c*=AH3_(ARcpH1(max(AH1_(1.0/32768.0),AH1_(1.0)-AMax3H1(c.r,c.g,c.b))));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrSrtmHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB){ + AH2 rcp=ARcpH2(AMax3H2(cR,cG,cB)+AH2_(1.0));cR*=rcp;cG*=rcp;cB*=rcp;} + void FsrSrtmInvHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB){ + AH2 rcp=ARcpH2(max(AH2_(1.0/32768.0),AH2_(1.0)-AMax3H2(cR,cG,cB)));cR*=rcp;cG*=rcp;cB*=rcp;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [TEPD] TEMPORAL ENERGY PRESERVING DITHER +// +//------------------------------------------------------------------------------------------------------------------------------ +// Temporally energy preserving dithered {0 to 1} linear to gamma 2.0 conversion. +// Gamma 2.0 is used so that the conversion back to linear is just to square the color. +// The conversion comes in 8-bit and 10-bit modes, designed for output to 8-bit UNORM or 10:10:10:2 respectively. +// Given good non-biased temporal blue noise as dither input, +// the output dither will temporally conserve energy. +// This is done by choosing the linear nearest step point instead of perceptual nearest. +// See code below for details. +//------------------------------------------------------------------------------------------------------------------------------ +// DX SPEC RULES FOR FLOAT->UNORM 8-BIT CONVERSION +// =============================================== +// - Output is 'uint(floor(saturate(n)*255.0+0.5))'. +// - Thus rounding is to nearest. +// - NaN gets converted to zero. +// - INF is clamped to {0.0 to 1.0}. +//============================================================================================================================== +#if defined(A_GPU) + // Hand tuned integer position to dither value, with more values than simple checkerboard. + // Only 32-bit has enough precision for this compddation. + // Output is {0 to <1}. + AF1 FsrTepdDitF(AU2 p,AU1 f){ + AF1 x=AF1_(p.x+f); + AF1 y=AF1_(p.y); + // The 1.61803 golden ratio. + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + // Number designed to provide a good visual pattern. + AF1 b=AF1_(1.0/3.69); + x=x*a+(y*b); + return AFractF1(x);} +//------------------------------------------------------------------------------------------------------------------------------ + // This version is 8-bit gamma 2.0. + // The 'c' input is {0 to 1}. + // Output is {0 to 1} ready for image store. + void FsrTepdC8F(inout AF3 c,AF1 dit){ + AF3 n=sqrt(c); + n=floor(n*AF3_(255.0))*AF3_(1.0/255.0); + AF3 a=n*n; + AF3 b=n+AF3_(1.0/255.0);b=b*b; + // Ratio of 'a' to 'b' required to produce 'c'. + // APrxLoRcpF1() won't work here (at least for very high dynamic ranges). + // APrxMedRcpF1() is an IADD,FMA,MUL. + AF3 r=(c-b)*APrxMedRcpF3(a-b); + // Use the ratio as a cutoff to choose 'a' or 'b'. + // AGtZeroF1() is a MUL. + c=ASatF3(n+AGtZeroF3(AF3_(dit)-r)*AF3_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + // This version is 10-bit gamma 2.0. + // The 'c' input is {0 to 1}. + // Output is {0 to 1} ready for image store. + void FsrTepdC10F(inout AF3 c,AF1 dit){ + AF3 n=sqrt(c); + n=floor(n*AF3_(1023.0))*AF3_(1.0/1023.0); + AF3 a=n*n; + AF3 b=n+AF3_(1.0/1023.0);b=b*b; + AF3 r=(c-b)*APrxMedRcpF3(a-b); + c=ASatF3(n+AGtZeroF3(AF3_(dit)-r)*AF3_(1.0/1023.0));} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + AH1 FsrTepdDitH(AU2 p,AU1 f){ + AF1 x=AF1_(p.x+f); + AF1 y=AF1_(p.y); + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + AF1 b=AF1_(1.0/3.69); + x=x*a+(y*b); + return AH1(AFractF1(x));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC8H(inout AH3 c,AH1 dit){ + AH3 n=sqrt(c); + n=floor(n*AH3_(255.0))*AH3_(1.0/255.0); + AH3 a=n*n; + AH3 b=n+AH3_(1.0/255.0);b=b*b; + AH3 r=(c-b)*APrxMedRcpH3(a-b); + c=ASatH3(n+AGtZeroH3(AH3_(dit)-r)*AH3_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC10H(inout AH3 c,AH1 dit){ + AH3 n=sqrt(c); + n=floor(n*AH3_(1023.0))*AH3_(1.0/1023.0); + AH3 a=n*n; + AH3 b=n+AH3_(1.0/1023.0);b=b*b; + AH3 r=(c-b)*APrxMedRcpH3(a-b); + c=ASatH3(n+AGtZeroH3(AH3_(dit)-r)*AH3_(1.0/1023.0));} +//============================================================================================================================== + // This computes dither for positions 'p' and 'p+{8,0}'. + AH2 FsrTepdDitHx2(AU2 p,AU1 f){ + AF2 x; + x.x=AF1_(p.x+f); + x.y=x.x+AF1_(8.0); + AF1 y=AF1_(p.y); + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + AF1 b=AF1_(1.0/3.69); + x=x*AF2_(a)+AF2_(y*b); + return AH2(AFractF2(x));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC8Hx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 dit){ + AH2 nR=sqrt(cR); + AH2 nG=sqrt(cG); + AH2 nB=sqrt(cB); + nR=floor(nR*AH2_(255.0))*AH2_(1.0/255.0); + nG=floor(nG*AH2_(255.0))*AH2_(1.0/255.0); + nB=floor(nB*AH2_(255.0))*AH2_(1.0/255.0); + AH2 aR=nR*nR; + AH2 aG=nG*nG; + AH2 aB=nB*nB; + AH2 bR=nR+AH2_(1.0/255.0);bR=bR*bR; + AH2 bG=nG+AH2_(1.0/255.0);bG=bG*bG; + AH2 bB=nB+AH2_(1.0/255.0);bB=bB*bB; + AH2 rR=(cR-bR)*APrxMedRcpH2(aR-bR); + AH2 rG=(cG-bG)*APrxMedRcpH2(aG-bG); + AH2 rB=(cB-bB)*APrxMedRcpH2(aB-bB); + cR=ASatH2(nR+AGtZeroH2(dit-rR)*AH2_(1.0/255.0)); + cG=ASatH2(nG+AGtZeroH2(dit-rG)*AH2_(1.0/255.0)); + cB=ASatH2(nB+AGtZeroH2(dit-rB)*AH2_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC10Hx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 dit){ + AH2 nR=sqrt(cR); + AH2 nG=sqrt(cG); + AH2 nB=sqrt(cB); + nR=floor(nR*AH2_(1023.0))*AH2_(1.0/1023.0); + nG=floor(nG*AH2_(1023.0))*AH2_(1.0/1023.0); + nB=floor(nB*AH2_(1023.0))*AH2_(1.0/1023.0); + AH2 aR=nR*nR; + AH2 aG=nG*nG; + AH2 aB=nB*nB; + AH2 bR=nR+AH2_(1.0/1023.0);bR=bR*bR; + AH2 bG=nG+AH2_(1.0/1023.0);bG=bG*bG; + AH2 bB=nB+AH2_(1.0/1023.0);bB=bB*bB; + AH2 rR=(cR-bR)*APrxMedRcpH2(aR-bR); + AH2 rG=(cG-bG)*APrxMedRcpH2(aG-bG); + AH2 rB=(cB-bB)*APrxMedRcpH2(aB-bB); + cR=ASatH2(nR+AGtZeroH2(dit-rR)*AH2_(1.0/1023.0)); + cG=ASatH2(nG+AGtZeroH2(dit-rG)*AH2_(1.0/1023.0)); + cB=ASatH2(nB+AGtZeroH2(dit-rB)*AH2_(1.0/1023.0));} +#endif + + +float insideBox(vec2 v) { + vec2 s = step(bLeft, v) - step(tRight, v); + return s.x * s.y; +} + +AF2 translateDest(AF2 pos) { + AF2 translatedPos = AF2(pos.x, pos.y); + translatedPos.x = dstX1 < dstX0 ? dstX1 - translatedPos.x : translatedPos.x; + translatedPos.y = dstY0 < dstY1 ? dstY1 + dstY0 - translatedPos.y - 1 : translatedPos.y; + return translatedPos; +} + +void CurrFilter(AU2 pos) +{ + if((insideBox(vec2(pos.x, pos.y))) == 0) { + imageStore(imgOutput, ASU2(pos.x, pos.y), AF4(0,0,0,1)); + return; + } + AF3 c; + FsrEasuF(c, AU2(pos.x - bLeft.x, pos.y - bLeft.y), con0, con1, con2, con3); + imageStore(imgOutput, ASU2(translateDest(pos)), AF4(c, 1)); +} + +void main() { + srcW = abs(srcX1 - srcX0); + srcH = abs(srcY1 - srcY0); + dstW = abs(dstX1 - dstX0); + dstH = abs(dstY1 - dstY0); + + AU2 gxy = ARmp8x8(gl_LocalInvocationID.x) + AU2(gl_WorkGroupID.x << 4u, gl_WorkGroupID.y << 4u); + + setBounds(vec2(dstX0 < dstX1 ? dstX0 : dstX1, dstY0 < dstY1 ? dstY0 : dstY1), + vec2(dstX1 > dstX0 ? dstX1 : dstX0, dstY1 > dstY0 ? dstY1 : dstY0)); + + // Upscaling + FsrEasuCon(con0, con1, con2, con3, + srcW, srcH, // Viewport size (top left aligned) in the input image which is to be scaled. + srcW, srcH, // The size of the input image. + dstW, dstH); // The output resolution. + + CurrFilter(gxy); + gxy.x += 8u; + CurrFilter(gxy); + gxy.y += 8u; + CurrFilter(gxy); + gxy.x -= 8u; + CurrFilter(gxy); +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrScaling.spv b/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrScaling.spv new file mode 100644 index 00000000..c15b72ec Binary files /dev/null and b/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrScaling.spv differ diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrSharpening.glsl b/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrSharpening.glsl new file mode 100644 index 00000000..785bc0c8 --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrSharpening.glsl @@ -0,0 +1,3904 @@ +// Sharpening +#version 430 core +layout (local_size_x = 64) in; +layout( rgba8, binding = 0, set = 3) uniform image2D imgOutput; +layout( binding = 2 ) uniform invResolution +{ + vec2 invResolution_data; +}; +layout( binding = 3 ) uniform outvResolution +{ + vec2 outvResolution_data; +}; +layout( binding = 1, set = 2) uniform sampler2D source; +layout( binding = 4 ) uniform sharpening +{ + float sharpening_data; +}; + +#define A_GPU 1 +#define A_GLSL 1 +//============================================================================================================================== +// +// [A] SHADER PORTABILITY 1.20210629 +// +//============================================================================================================================== +// FidelityFX Super Resolution Sample +// +// Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved. +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files(the "Software"), to deal +// in the Software without restriction, including without limitation the rights +// to use, copy, modify, merge, publish, distribute, sublicense, and / or sell +// copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions : +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN +// THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +// MIT LICENSE +// =========== +// Copyright (c) 2014 Michal Drobot (for concepts used in "FLOAT APPROXIMATIONS"). +// ----------- +// Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation +// files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, +// modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the +// Software is furnished to do so, subject to the following conditions: +// ----------- +// The above copyright notice and this permission notice shall be included in all copies or substantial portions of the +// Software. +// ----------- +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE +// WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR +// COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, +// ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +// ABOUT +// ===== +// Common central point for high-level shading language and C portability for various shader headers. +//------------------------------------------------------------------------------------------------------------------------------ +// DEFINES +// ======= +// A_CPU ..... Include the CPU related code. +// A_GPU ..... Include the GPU related code. +// A_GLSL .... Using GLSL. +// A_HLSL .... Using HLSL. +// A_HLSL_6_2 Using HLSL 6.2 with new 'uint16_t' and related types (requires '-enable-16bit-types'). +// A_NO_16_BIT_CAST Don't use instructions that are not availabe in SPIR-V (needed for running A_HLSL_6_2 on Vulkan) +// A_GCC ..... Using a GCC compatible compiler (else assume MSVC compatible compiler by default). +// ======= +// A_BYTE .... Support 8-bit integer. +// A_HALF .... Support 16-bit integer and floating point. +// A_LONG .... Support 64-bit integer. +// A_DUBL .... Support 64-bit floating point. +// ======= +// A_WAVE .... Support wave-wide operations. +//------------------------------------------------------------------------------------------------------------------------------ +// To get #include "ffx_a.h" working in GLSL use '#extension GL_GOOGLE_include_directive:require'. +//------------------------------------------------------------------------------------------------------------------------------ +// SIMPLIFIED TYPE SYSTEM +// ====================== +// - All ints will be unsigned with exception of when signed is required. +// - Type naming simplified and shortened "A<#components>", +// - H = 16-bit float (half) +// - F = 32-bit float (float) +// - D = 64-bit float (double) +// - P = 1-bit integer (predicate, not using bool because 'B' is used for byte) +// - B = 8-bit integer (byte) +// - W = 16-bit integer (word) +// - U = 32-bit integer (unsigned) +// - L = 64-bit integer (long) +// - Using "AS<#components>" for signed when required. +//------------------------------------------------------------------------------------------------------------------------------ +// TODO +// ==== +// - Make sure 'ALerp*(a,b,m)' does 'b*m+(-a*m+a)' (2 ops). +//------------------------------------------------------------------------------------------------------------------------------ +// CHANGE LOG +// ========== +// 20200914 - Expanded wave ops and prx code. +// 20200713 - Added [ZOL] section, fixed serious bugs in sRGB and Rec.709 color conversion code, etc. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// COMMON +//============================================================================================================================== +#define A_2PI 6.28318530718 +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// CPU +// +// +//============================================================================================================================== +#ifdef A_CPU + // Supporting user defined overrides. + #ifndef A_RESTRICT + #define A_RESTRICT __restrict + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifndef A_STATIC + #define A_STATIC static + #endif +//------------------------------------------------------------------------------------------------------------------------------ + // Same types across CPU and GPU. + // Predicate uses 32-bit integer (C friendly bool). + typedef uint32_t AP1; + typedef float AF1; + typedef double AD1; + typedef uint8_t AB1; + typedef uint16_t AW1; + typedef uint32_t AU1; + typedef uint64_t AL1; + typedef int8_t ASB1; + typedef int16_t ASW1; + typedef int32_t ASU1; + typedef int64_t ASL1; +//------------------------------------------------------------------------------------------------------------------------------ + #define AD1_(a) ((AD1)(a)) + #define AF1_(a) ((AF1)(a)) + #define AL1_(a) ((AL1)(a)) + #define AU1_(a) ((AU1)(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ASL1_(a) ((ASL1)(a)) + #define ASU1_(a) ((ASU1)(a)) +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AU1 AU1_AF1(AF1 a){union{AF1 f;AU1 u;}bits;bits.f=a;return bits.u;} +//------------------------------------------------------------------------------------------------------------------------------ + #define A_TRUE 1 + #define A_FALSE 0 +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// CPU/GPU PORTING +// +//------------------------------------------------------------------------------------------------------------------------------ +// Get CPU and GPU to share all setup code, without duplicate code paths. +// This uses a lower-case prefix for special vector constructs. +// - In C restrict pointers are used. +// - In the shading language, in/inout/out arguments are used. +// This depends on the ability to access a vector value in both languages via array syntax (aka color[2]). +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR ARGUMENT/RETURN/INITIALIZATION PORTABILITY +//============================================================================================================================== + #define retAD2 AD1 *A_RESTRICT + #define retAD3 AD1 *A_RESTRICT + #define retAD4 AD1 *A_RESTRICT + #define retAF2 AF1 *A_RESTRICT + #define retAF3 AF1 *A_RESTRICT + #define retAF4 AF1 *A_RESTRICT + #define retAL2 AL1 *A_RESTRICT + #define retAL3 AL1 *A_RESTRICT + #define retAL4 AL1 *A_RESTRICT + #define retAU2 AU1 *A_RESTRICT + #define retAU3 AU1 *A_RESTRICT + #define retAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define inAD2 AD1 *A_RESTRICT + #define inAD3 AD1 *A_RESTRICT + #define inAD4 AD1 *A_RESTRICT + #define inAF2 AF1 *A_RESTRICT + #define inAF3 AF1 *A_RESTRICT + #define inAF4 AF1 *A_RESTRICT + #define inAL2 AL1 *A_RESTRICT + #define inAL3 AL1 *A_RESTRICT + #define inAL4 AL1 *A_RESTRICT + #define inAU2 AU1 *A_RESTRICT + #define inAU3 AU1 *A_RESTRICT + #define inAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define inoutAD2 AD1 *A_RESTRICT + #define inoutAD3 AD1 *A_RESTRICT + #define inoutAD4 AD1 *A_RESTRICT + #define inoutAF2 AF1 *A_RESTRICT + #define inoutAF3 AF1 *A_RESTRICT + #define inoutAF4 AF1 *A_RESTRICT + #define inoutAL2 AL1 *A_RESTRICT + #define inoutAL3 AL1 *A_RESTRICT + #define inoutAL4 AL1 *A_RESTRICT + #define inoutAU2 AU1 *A_RESTRICT + #define inoutAU3 AU1 *A_RESTRICT + #define inoutAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define outAD2 AD1 *A_RESTRICT + #define outAD3 AD1 *A_RESTRICT + #define outAD4 AD1 *A_RESTRICT + #define outAF2 AF1 *A_RESTRICT + #define outAF3 AF1 *A_RESTRICT + #define outAF4 AF1 *A_RESTRICT + #define outAL2 AL1 *A_RESTRICT + #define outAL3 AL1 *A_RESTRICT + #define outAL4 AL1 *A_RESTRICT + #define outAU2 AU1 *A_RESTRICT + #define outAU3 AU1 *A_RESTRICT + #define outAU4 AU1 *A_RESTRICT +//------------------------------------------------------------------------------------------------------------------------------ + #define varAD2(x) AD1 x[2] + #define varAD3(x) AD1 x[3] + #define varAD4(x) AD1 x[4] + #define varAF2(x) AF1 x[2] + #define varAF3(x) AF1 x[3] + #define varAF4(x) AF1 x[4] + #define varAL2(x) AL1 x[2] + #define varAL3(x) AL1 x[3] + #define varAL4(x) AL1 x[4] + #define varAU2(x) AU1 x[2] + #define varAU3(x) AU1 x[3] + #define varAU4(x) AU1 x[4] +//------------------------------------------------------------------------------------------------------------------------------ + #define initAD2(x,y) {x,y} + #define initAD3(x,y,z) {x,y,z} + #define initAD4(x,y,z,w) {x,y,z,w} + #define initAF2(x,y) {x,y} + #define initAF3(x,y,z) {x,y,z} + #define initAF4(x,y,z,w) {x,y,z,w} + #define initAL2(x,y) {x,y} + #define initAL3(x,y,z) {x,y,z} + #define initAL4(x,y,z,w) {x,y,z,w} + #define initAU2(x,y) {x,y} + #define initAU3(x,y,z) {x,y,z} + #define initAU4(x,y,z,w) {x,y,z,w} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS +//------------------------------------------------------------------------------------------------------------------------------ +// TODO +// ==== +// - Replace transcendentals with manual versions. +//============================================================================================================================== + #ifdef A_GCC + A_STATIC AD1 AAbsD1(AD1 a){return __builtin_fabs(a);} + A_STATIC AF1 AAbsF1(AF1 a){return __builtin_fabsf(a);} + A_STATIC AU1 AAbsSU1(AU1 a){return AU1_(__builtin_abs(ASU1_(a)));} + A_STATIC AL1 AAbsSL1(AL1 a){return AL1_(__builtin_llabs(ASL1_(a)));} + #else + A_STATIC AD1 AAbsD1(AD1 a){return fabs(a);} + A_STATIC AF1 AAbsF1(AF1 a){return fabsf(a);} + A_STATIC AU1 AAbsSU1(AU1 a){return AU1_(abs(ASU1_(a)));} + A_STATIC AL1 AAbsSL1(AL1 a){return AL1_(labs((long)ASL1_(a)));} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ACosD1(AD1 a){return __builtin_cos(a);} + A_STATIC AF1 ACosF1(AF1 a){return __builtin_cosf(a);} + #else + A_STATIC AD1 ACosD1(AD1 a){return cos(a);} + A_STATIC AF1 ACosF1(AF1 a){return cosf(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ADotD2(inAD2 a,inAD2 b){return a[0]*b[0]+a[1]*b[1];} + A_STATIC AD1 ADotD3(inAD3 a,inAD3 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2];} + A_STATIC AD1 ADotD4(inAD4 a,inAD4 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2]+a[3]*b[3];} + A_STATIC AF1 ADotF2(inAF2 a,inAF2 b){return a[0]*b[0]+a[1]*b[1];} + A_STATIC AF1 ADotF3(inAF3 a,inAF3 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2];} + A_STATIC AF1 ADotF4(inAF4 a,inAF4 b){return a[0]*b[0]+a[1]*b[1]+a[2]*b[2]+a[3]*b[3];} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 AExp2D1(AD1 a){return __builtin_exp2(a);} + A_STATIC AF1 AExp2F1(AF1 a){return __builtin_exp2f(a);} + #else + A_STATIC AD1 AExp2D1(AD1 a){return exp2(a);} + A_STATIC AF1 AExp2F1(AF1 a){return exp2f(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 AFloorD1(AD1 a){return __builtin_floor(a);} + A_STATIC AF1 AFloorF1(AF1 a){return __builtin_floorf(a);} + #else + A_STATIC AD1 AFloorD1(AD1 a){return floor(a);} + A_STATIC AF1 AFloorF1(AF1 a){return floorf(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ALerpD1(AD1 a,AD1 b,AD1 c){return b*c+(-a*c+a);} + A_STATIC AF1 ALerpF1(AF1 a,AF1 b,AF1 c){return b*c+(-a*c+a);} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ALog2D1(AD1 a){return __builtin_log2(a);} + A_STATIC AF1 ALog2F1(AF1 a){return __builtin_log2f(a);} + #else + A_STATIC AD1 ALog2D1(AD1 a){return log2(a);} + A_STATIC AF1 ALog2F1(AF1 a){return log2f(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 AMaxD1(AD1 a,AD1 b){return a>b?a:b;} + A_STATIC AF1 AMaxF1(AF1 a,AF1 b){return a>b?a:b;} + A_STATIC AL1 AMaxL1(AL1 a,AL1 b){return a>b?a:b;} + A_STATIC AU1 AMaxU1(AU1 a,AU1 b){return a>b?a:b;} +//------------------------------------------------------------------------------------------------------------------------------ + // These follow the convention that A integer types don't have signage, until they are operated on. + A_STATIC AL1 AMaxSL1(AL1 a,AL1 b){return (ASL1_(a)>ASL1_(b))?a:b;} + A_STATIC AU1 AMaxSU1(AU1 a,AU1 b){return (ASU1_(a)>ASU1_(b))?a:b;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 AMinD1(AD1 a,AD1 b){return a>ASL1_(b));} + A_STATIC AU1 AShrSU1(AU1 a,AU1 b){return AU1_(ASU1_(a)>>ASU1_(b));} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ASinD1(AD1 a){return __builtin_sin(a);} + A_STATIC AF1 ASinF1(AF1 a){return __builtin_sinf(a);} + #else + A_STATIC AD1 ASinD1(AD1 a){return sin(a);} + A_STATIC AF1 ASinF1(AF1 a){return sinf(a);} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_GCC + A_STATIC AD1 ASqrtD1(AD1 a){return __builtin_sqrt(a);} + A_STATIC AF1 ASqrtF1(AF1 a){return __builtin_sqrtf(a);} + #else + A_STATIC AD1 ASqrtD1(AD1 a){return sqrt(a);} + A_STATIC AF1 ASqrtF1(AF1 a){return sqrtf(a);} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS - DEPENDENT +//============================================================================================================================== + A_STATIC AD1 AClampD1(AD1 x,AD1 n,AD1 m){return AMaxD1(n,AMinD1(x,m));} + A_STATIC AF1 AClampF1(AF1 x,AF1 n,AF1 m){return AMaxF1(n,AMinF1(x,m));} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 AFractD1(AD1 a){return a-AFloorD1(a);} + A_STATIC AF1 AFractF1(AF1 a){return a-AFloorF1(a);} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 APowD1(AD1 a,AD1 b){return AExp2D1(b*ALog2D1(a));} + A_STATIC AF1 APowF1(AF1 a,AF1 b){return AExp2F1(b*ALog2F1(a));} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ARsqD1(AD1 a){return ARcpD1(ASqrtD1(a));} + A_STATIC AF1 ARsqF1(AF1 a){return ARcpF1(ASqrtF1(a));} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC AD1 ASatD1(AD1 a){return AMinD1(1.0,AMaxD1(0.0,a));} + A_STATIC AF1 ASatF1(AF1 a){return AMinF1(1.0f,AMaxF1(0.0f,a));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR OPS +//------------------------------------------------------------------------------------------------------------------------------ +// These are added as needed for production or prototyping, so not necessarily a complete set. +// They follow a convention of taking in a destination and also returning the destination value to increase utility. +//============================================================================================================================== + A_STATIC retAD2 opAAbsD2(outAD2 d,inAD2 a){d[0]=AAbsD1(a[0]);d[1]=AAbsD1(a[1]);return d;} + A_STATIC retAD3 opAAbsD3(outAD3 d,inAD3 a){d[0]=AAbsD1(a[0]);d[1]=AAbsD1(a[1]);d[2]=AAbsD1(a[2]);return d;} + A_STATIC retAD4 opAAbsD4(outAD4 d,inAD4 a){d[0]=AAbsD1(a[0]);d[1]=AAbsD1(a[1]);d[2]=AAbsD1(a[2]);d[3]=AAbsD1(a[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAAbsF2(outAF2 d,inAF2 a){d[0]=AAbsF1(a[0]);d[1]=AAbsF1(a[1]);return d;} + A_STATIC retAF3 opAAbsF3(outAF3 d,inAF3 a){d[0]=AAbsF1(a[0]);d[1]=AAbsF1(a[1]);d[2]=AAbsF1(a[2]);return d;} + A_STATIC retAF4 opAAbsF4(outAF4 d,inAF4 a){d[0]=AAbsF1(a[0]);d[1]=AAbsF1(a[1]);d[2]=AAbsF1(a[2]);d[3]=AAbsF1(a[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAAddD2(outAD2 d,inAD2 a,inAD2 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];return d;} + A_STATIC retAD3 opAAddD3(outAD3 d,inAD3 a,inAD3 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];return d;} + A_STATIC retAD4 opAAddD4(outAD4 d,inAD4 a,inAD4 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];d[3]=a[3]+b[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAAddF2(outAF2 d,inAF2 a,inAF2 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];return d;} + A_STATIC retAF3 opAAddF3(outAF3 d,inAF3 a,inAF3 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];return d;} + A_STATIC retAF4 opAAddF4(outAF4 d,inAF4 a,inAF4 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];d[3]=a[3]+b[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opAAddOneD2(outAD2 d,inAD2 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;return d;} + A_STATIC retAD3 opAAddOneD3(outAD3 d,inAD3 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;return d;} + A_STATIC retAD4 opAAddOneD4(outAD4 d,inAD4 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;d[3]=a[3]+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAAddOneF2(outAF2 d,inAF2 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;return d;} + A_STATIC retAF3 opAAddOneF3(outAF3 d,inAF3 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;return d;} + A_STATIC retAF4 opAAddOneF4(outAF4 d,inAF4 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;d[3]=a[3]+b;return d;} +//============================================================================================================================== + A_STATIC retAD2 opACpyD2(outAD2 d,inAD2 a){d[0]=a[0];d[1]=a[1];return d;} + A_STATIC retAD3 opACpyD3(outAD3 d,inAD3 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];return d;} + A_STATIC retAD4 opACpyD4(outAD4 d,inAD4 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];d[3]=a[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opACpyF2(outAF2 d,inAF2 a){d[0]=a[0];d[1]=a[1];return d;} + A_STATIC retAF3 opACpyF3(outAF3 d,inAF3 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];return d;} + A_STATIC retAF4 opACpyF4(outAF4 d,inAF4 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];d[3]=a[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opALerpD2(outAD2 d,inAD2 a,inAD2 b,inAD2 c){d[0]=ALerpD1(a[0],b[0],c[0]);d[1]=ALerpD1(a[1],b[1],c[1]);return d;} + A_STATIC retAD3 opALerpD3(outAD3 d,inAD3 a,inAD3 b,inAD3 c){d[0]=ALerpD1(a[0],b[0],c[0]);d[1]=ALerpD1(a[1],b[1],c[1]);d[2]=ALerpD1(a[2],b[2],c[2]);return d;} + A_STATIC retAD4 opALerpD4(outAD4 d,inAD4 a,inAD4 b,inAD4 c){d[0]=ALerpD1(a[0],b[0],c[0]);d[1]=ALerpD1(a[1],b[1],c[1]);d[2]=ALerpD1(a[2],b[2],c[2]);d[3]=ALerpD1(a[3],b[3],c[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opALerpF2(outAF2 d,inAF2 a,inAF2 b,inAF2 c){d[0]=ALerpF1(a[0],b[0],c[0]);d[1]=ALerpF1(a[1],b[1],c[1]);return d;} + A_STATIC retAF3 opALerpF3(outAF3 d,inAF3 a,inAF3 b,inAF3 c){d[0]=ALerpF1(a[0],b[0],c[0]);d[1]=ALerpF1(a[1],b[1],c[1]);d[2]=ALerpF1(a[2],b[2],c[2]);return d;} + A_STATIC retAF4 opALerpF4(outAF4 d,inAF4 a,inAF4 b,inAF4 c){d[0]=ALerpF1(a[0],b[0],c[0]);d[1]=ALerpF1(a[1],b[1],c[1]);d[2]=ALerpF1(a[2],b[2],c[2]);d[3]=ALerpF1(a[3],b[3],c[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opALerpOneD2(outAD2 d,inAD2 a,inAD2 b,AD1 c){d[0]=ALerpD1(a[0],b[0],c);d[1]=ALerpD1(a[1],b[1],c);return d;} + A_STATIC retAD3 opALerpOneD3(outAD3 d,inAD3 a,inAD3 b,AD1 c){d[0]=ALerpD1(a[0],b[0],c);d[1]=ALerpD1(a[1],b[1],c);d[2]=ALerpD1(a[2],b[2],c);return d;} + A_STATIC retAD4 opALerpOneD4(outAD4 d,inAD4 a,inAD4 b,AD1 c){d[0]=ALerpD1(a[0],b[0],c);d[1]=ALerpD1(a[1],b[1],c);d[2]=ALerpD1(a[2],b[2],c);d[3]=ALerpD1(a[3],b[3],c);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opALerpOneF2(outAF2 d,inAF2 a,inAF2 b,AF1 c){d[0]=ALerpF1(a[0],b[0],c);d[1]=ALerpF1(a[1],b[1],c);return d;} + A_STATIC retAF3 opALerpOneF3(outAF3 d,inAF3 a,inAF3 b,AF1 c){d[0]=ALerpF1(a[0],b[0],c);d[1]=ALerpF1(a[1],b[1],c);d[2]=ALerpF1(a[2],b[2],c);return d;} + A_STATIC retAF4 opALerpOneF4(outAF4 d,inAF4 a,inAF4 b,AF1 c){d[0]=ALerpF1(a[0],b[0],c);d[1]=ALerpF1(a[1],b[1],c);d[2]=ALerpF1(a[2],b[2],c);d[3]=ALerpF1(a[3],b[3],c);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMaxD2(outAD2 d,inAD2 a,inAD2 b){d[0]=AMaxD1(a[0],b[0]);d[1]=AMaxD1(a[1],b[1]);return d;} + A_STATIC retAD3 opAMaxD3(outAD3 d,inAD3 a,inAD3 b){d[0]=AMaxD1(a[0],b[0]);d[1]=AMaxD1(a[1],b[1]);d[2]=AMaxD1(a[2],b[2]);return d;} + A_STATIC retAD4 opAMaxD4(outAD4 d,inAD4 a,inAD4 b){d[0]=AMaxD1(a[0],b[0]);d[1]=AMaxD1(a[1],b[1]);d[2]=AMaxD1(a[2],b[2]);d[3]=AMaxD1(a[3],b[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMaxF2(outAF2 d,inAF2 a,inAF2 b){d[0]=AMaxF1(a[0],b[0]);d[1]=AMaxF1(a[1],b[1]);return d;} + A_STATIC retAF3 opAMaxF3(outAF3 d,inAF3 a,inAF3 b){d[0]=AMaxF1(a[0],b[0]);d[1]=AMaxF1(a[1],b[1]);d[2]=AMaxF1(a[2],b[2]);return d;} + A_STATIC retAF4 opAMaxF4(outAF4 d,inAF4 a,inAF4 b){d[0]=AMaxF1(a[0],b[0]);d[1]=AMaxF1(a[1],b[1]);d[2]=AMaxF1(a[2],b[2]);d[3]=AMaxF1(a[3],b[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMinD2(outAD2 d,inAD2 a,inAD2 b){d[0]=AMinD1(a[0],b[0]);d[1]=AMinD1(a[1],b[1]);return d;} + A_STATIC retAD3 opAMinD3(outAD3 d,inAD3 a,inAD3 b){d[0]=AMinD1(a[0],b[0]);d[1]=AMinD1(a[1],b[1]);d[2]=AMinD1(a[2],b[2]);return d;} + A_STATIC retAD4 opAMinD4(outAD4 d,inAD4 a,inAD4 b){d[0]=AMinD1(a[0],b[0]);d[1]=AMinD1(a[1],b[1]);d[2]=AMinD1(a[2],b[2]);d[3]=AMinD1(a[3],b[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMinF2(outAF2 d,inAF2 a,inAF2 b){d[0]=AMinF1(a[0],b[0]);d[1]=AMinF1(a[1],b[1]);return d;} + A_STATIC retAF3 opAMinF3(outAF3 d,inAF3 a,inAF3 b){d[0]=AMinF1(a[0],b[0]);d[1]=AMinF1(a[1],b[1]);d[2]=AMinF1(a[2],b[2]);return d;} + A_STATIC retAF4 opAMinF4(outAF4 d,inAF4 a,inAF4 b){d[0]=AMinF1(a[0],b[0]);d[1]=AMinF1(a[1],b[1]);d[2]=AMinF1(a[2],b[2]);d[3]=AMinF1(a[3],b[3]);return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMulD2(outAD2 d,inAD2 a,inAD2 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];return d;} + A_STATIC retAD3 opAMulD3(outAD3 d,inAD3 a,inAD3 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];return d;} + A_STATIC retAD4 opAMulD4(outAD4 d,inAD4 a,inAD4 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];d[3]=a[3]*b[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMulF2(outAF2 d,inAF2 a,inAF2 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];return d;} + A_STATIC retAF3 opAMulF3(outAF3 d,inAF3 a,inAF3 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];return d;} + A_STATIC retAF4 opAMulF4(outAF4 d,inAF4 a,inAF4 b){d[0]=a[0]*b[0];d[1]=a[1]*b[1];d[2]=a[2]*b[2];d[3]=a[3]*b[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opAMulOneD2(outAD2 d,inAD2 a,AD1 b){d[0]=a[0]*b;d[1]=a[1]*b;return d;} + A_STATIC retAD3 opAMulOneD3(outAD3 d,inAD3 a,AD1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;return d;} + A_STATIC retAD4 opAMulOneD4(outAD4 d,inAD4 a,AD1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;d[3]=a[3]*b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAMulOneF2(outAF2 d,inAF2 a,AF1 b){d[0]=a[0]*b;d[1]=a[1]*b;return d;} + A_STATIC retAF3 opAMulOneF3(outAF3 d,inAF3 a,AF1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;return d;} + A_STATIC retAF4 opAMulOneF4(outAF4 d,inAF4 a,AF1 b){d[0]=a[0]*b;d[1]=a[1]*b;d[2]=a[2]*b;d[3]=a[3]*b;return d;} +//============================================================================================================================== + A_STATIC retAD2 opANegD2(outAD2 d,inAD2 a){d[0]=-a[0];d[1]=-a[1];return d;} + A_STATIC retAD3 opANegD3(outAD3 d,inAD3 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];return d;} + A_STATIC retAD4 opANegD4(outAD4 d,inAD4 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];d[3]=-a[3];return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opANegF2(outAF2 d,inAF2 a){d[0]=-a[0];d[1]=-a[1];return d;} + A_STATIC retAF3 opANegF3(outAF3 d,inAF3 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];return d;} + A_STATIC retAF4 opANegF4(outAF4 d,inAF4 a){d[0]=-a[0];d[1]=-a[1];d[2]=-a[2];d[3]=-a[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opARcpD2(outAD2 d,inAD2 a){d[0]=ARcpD1(a[0]);d[1]=ARcpD1(a[1]);return d;} + A_STATIC retAD3 opARcpD3(outAD3 d,inAD3 a){d[0]=ARcpD1(a[0]);d[1]=ARcpD1(a[1]);d[2]=ARcpD1(a[2]);return d;} + A_STATIC retAD4 opARcpD4(outAD4 d,inAD4 a){d[0]=ARcpD1(a[0]);d[1]=ARcpD1(a[1]);d[2]=ARcpD1(a[2]);d[3]=ARcpD1(a[3]);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opARcpF2(outAF2 d,inAF2 a){d[0]=ARcpF1(a[0]);d[1]=ARcpF1(a[1]);return d;} + A_STATIC retAF3 opARcpF3(outAF3 d,inAF3 a){d[0]=ARcpF1(a[0]);d[1]=ARcpF1(a[1]);d[2]=ARcpF1(a[2]);return d;} + A_STATIC retAF4 opARcpF4(outAF4 d,inAF4 a){d[0]=ARcpF1(a[0]);d[1]=ARcpF1(a[1]);d[2]=ARcpF1(a[2]);d[3]=ARcpF1(a[3]);return d;} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HALF FLOAT PACKING +//============================================================================================================================== + // Convert float to half (in lower 16-bits of output). + // Same fast technique as documented here: ftp://ftp.fox-toolkit.org/pub/fasthalffloatconversion.pdf + // Supports denormals. + // Conversion rules are to make computations possibly "safer" on the GPU, + // -INF & -NaN -> -65504 + // +INF & +NaN -> +65504 + A_STATIC AU1 AU1_AH1_AF1(AF1 f){ + static AW1 base[512]={ + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000, + 0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0001,0x0002,0x0004,0x0008,0x0010,0x0020,0x0040,0x0080,0x0100, + 0x0200,0x0400,0x0800,0x0c00,0x1000,0x1400,0x1800,0x1c00,0x2000,0x2400,0x2800,0x2c00,0x3000,0x3400,0x3800,0x3c00, + 0x4000,0x4400,0x4800,0x4c00,0x5000,0x5400,0x5800,0x5c00,0x6000,0x6400,0x6800,0x6c00,0x7000,0x7400,0x7800,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff,0x7bff, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000, + 0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8001,0x8002,0x8004,0x8008,0x8010,0x8020,0x8040,0x8080,0x8100, + 0x8200,0x8400,0x8800,0x8c00,0x9000,0x9400,0x9800,0x9c00,0xa000,0xa400,0xa800,0xac00,0xb000,0xb400,0xb800,0xbc00, + 0xc000,0xc400,0xc800,0xcc00,0xd000,0xd400,0xd800,0xdc00,0xe000,0xe400,0xe800,0xec00,0xf000,0xf400,0xf800,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff, + 0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff,0xfbff}; + static AB1 shift[512]={ + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x17,0x16,0x15,0x14,0x13,0x12,0x11,0x10,0x0f, + 0x0e,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d, + 0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x17,0x16,0x15,0x14,0x13,0x12,0x11,0x10,0x0f, + 0x0e,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d, + 0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x0d,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18, + 0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18,0x18}; + union{AF1 f;AU1 u;}bits;bits.f=f;AU1 u=bits.u;AU1 i=u>>23;return (AU1)(base[i])+((u&0x7fffff)>>shift[i]);} +//------------------------------------------------------------------------------------------------------------------------------ + // Used to output packed constant. + A_STATIC AU1 AU1_AH2_AF2(inAF2 a){return AU1_AH1_AF1(a[0])+(AU1_AH1_AF1(a[1])<<16);} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// GLSL +// +// +//============================================================================================================================== +#if defined(A_GLSL) && defined(A_GPU) + #ifndef A_SKIP_EXT + #ifdef A_HALF + #extension GL_EXT_shader_16bit_storage:require + #extension GL_EXT_shader_explicit_arithmetic_types:require + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_LONG + #extension GL_ARB_gpu_shader_int64:require + #extension GL_NV_shader_atomic_int64:require + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_WAVE + #extension GL_KHR_shader_subgroup_arithmetic:require + #extension GL_KHR_shader_subgroup_ballot:require + #extension GL_KHR_shader_subgroup_quad:require + #extension GL_KHR_shader_subgroup_shuffle:require + #endif + #endif +//============================================================================================================================== + #define AP1 bool + #define AP2 bvec2 + #define AP3 bvec3 + #define AP4 bvec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF1 float + #define AF2 vec2 + #define AF3 vec3 + #define AF4 vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1 uint + #define AU2 uvec2 + #define AU3 uvec3 + #define AU4 uvec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASU1 int + #define ASU2 ivec2 + #define ASU3 ivec3 + #define ASU4 ivec4 +//============================================================================================================================== + #define AF1_AU1(x) uintBitsToFloat(AU1(x)) + #define AF2_AU2(x) uintBitsToFloat(AU2(x)) + #define AF3_AU3(x) uintBitsToFloat(AU3(x)) + #define AF4_AU4(x) uintBitsToFloat(AU4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AF1(x) floatBitsToUint(AF1(x)) + #define AU2_AF2(x) floatBitsToUint(AF2(x)) + #define AU3_AF3(x) floatBitsToUint(AF3(x)) + #define AU4_AF4(x) floatBitsToUint(AF4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH1_AF1_x(AF1 a){return packHalf2x16(AF2(a,0.0));} + #define AU1_AH1_AF1(a) AU1_AH1_AF1_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AH2_AF2 packHalf2x16 + #define AU1_AW2Unorm_AF2 packUnorm2x16 + #define AU1_AB4Unorm_AF4 packUnorm4x8 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF2_AH2_AU1 unpackHalf2x16 + #define AF2_AW2Unorm_AU1 unpackUnorm2x16 + #define AF4_AB4Unorm_AU1 unpackUnorm4x8 +//============================================================================================================================== + AF1 AF1_x(AF1 a){return AF1(a);} + AF2 AF2_x(AF1 a){return AF2(a,a);} + AF3 AF3_x(AF1 a){return AF3(a,a,a);} + AF4 AF4_x(AF1 a){return AF4(a,a,a,a);} + #define AF1_(a) AF1_x(AF1(a)) + #define AF2_(a) AF2_x(AF1(a)) + #define AF3_(a) AF3_x(AF1(a)) + #define AF4_(a) AF4_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_x(AU1 a){return AU1(a);} + AU2 AU2_x(AU1 a){return AU2(a,a);} + AU3 AU3_x(AU1 a){return AU3(a,a,a);} + AU4 AU4_x(AU1 a){return AU4(a,a,a,a);} + #define AU1_(a) AU1_x(AU1(a)) + #define AU2_(a) AU2_x(AU1(a)) + #define AU3_(a) AU3_x(AU1(a)) + #define AU4_(a) AU4_x(AU1(a)) +//============================================================================================================================== + AU1 AAbsSU1(AU1 a){return AU1(abs(ASU1(a)));} + AU2 AAbsSU2(AU2 a){return AU2(abs(ASU2(a)));} + AU3 AAbsSU3(AU3 a){return AU3(abs(ASU3(a)));} + AU4 AAbsSU4(AU4 a){return AU4(abs(ASU4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 ABfe(AU1 src,AU1 off,AU1 bits){return bitfieldExtract(src,ASU1(off),ASU1(bits));} + AU1 ABfi(AU1 src,AU1 ins,AU1 mask){return (ins&mask)|(src&(~mask));} + // Proxy for V_BFI_B32 where the 'mask' is set as 'bits', 'mask=(1<>ASU1(b));} + AU2 AShrSU2(AU2 a,AU2 b){return AU2(ASU2(a)>>ASU2(b));} + AU3 AShrSU3(AU3 a,AU3 b){return AU3(ASU3(a)>>ASU3(b));} + AU4 AShrSU4(AU4 a,AU4 b){return AU4(ASU4(a)>>ASU4(b));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL BYTE +//============================================================================================================================== + #ifdef A_BYTE + #define AB1 uint8_t + #define AB2 u8vec2 + #define AB3 u8vec3 + #define AB4 u8vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASB1 int8_t + #define ASB2 i8vec2 + #define ASB3 i8vec3 + #define ASB4 i8vec4 +//------------------------------------------------------------------------------------------------------------------------------ + AB1 AB1_x(AB1 a){return AB1(a);} + AB2 AB2_x(AB1 a){return AB2(a,a);} + AB3 AB3_x(AB1 a){return AB3(a,a,a);} + AB4 AB4_x(AB1 a){return AB4(a,a,a,a);} + #define AB1_(a) AB1_x(AB1(a)) + #define AB2_(a) AB2_x(AB1(a)) + #define AB3_(a) AB3_x(AB1(a)) + #define AB4_(a) AB4_x(AB1(a)) + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL HALF +//============================================================================================================================== + #ifdef A_HALF + #define AH1 float16_t + #define AH2 f16vec2 + #define AH3 f16vec3 + #define AH4 f16vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AW1 uint16_t + #define AW2 u16vec2 + #define AW3 u16vec3 + #define AW4 u16vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASW1 int16_t + #define ASW2 i16vec2 + #define ASW3 i16vec3 + #define ASW4 i16vec4 +//============================================================================================================================== + #define AH2_AU1(x) unpackFloat2x16(AU1(x)) + AH4 AH4_AU2_x(AU2 x){return AH4(unpackFloat2x16(x.x),unpackFloat2x16(x.y));} + #define AH4_AU2(x) AH4_AU2_x(AU2(x)) + #define AW2_AU1(x) unpackUint2x16(AU1(x)) + #define AW4_AU2(x) unpackUint4x16(pack64(AU2(x))) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AH2(x) packFloat2x16(AH2(x)) + AU2 AU2_AH4_x(AH4 x){return AU2(packFloat2x16(x.xy),packFloat2x16(x.zw));} + #define AU2_AH4(x) AU2_AH4_x(AH4(x)) + #define AU1_AW2(x) packUint2x16(AW2(x)) + #define AU2_AW4(x) unpack32(packUint4x16(AW4(x))) +//============================================================================================================================== + #define AW1_AH1(x) halfBitsToUint16(AH1(x)) + #define AW2_AH2(x) halfBitsToUint16(AH2(x)) + #define AW3_AH3(x) halfBitsToUint16(AH3(x)) + #define AW4_AH4(x) halfBitsToUint16(AH4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AH1_AW1(x) uint16BitsToHalf(AW1(x)) + #define AH2_AW2(x) uint16BitsToHalf(AW2(x)) + #define AH3_AW3(x) uint16BitsToHalf(AW3(x)) + #define AH4_AW4(x) uint16BitsToHalf(AW4(x)) +//============================================================================================================================== + AH1 AH1_x(AH1 a){return AH1(a);} + AH2 AH2_x(AH1 a){return AH2(a,a);} + AH3 AH3_x(AH1 a){return AH3(a,a,a);} + AH4 AH4_x(AH1 a){return AH4(a,a,a,a);} + #define AH1_(a) AH1_x(AH1(a)) + #define AH2_(a) AH2_x(AH1(a)) + #define AH3_(a) AH3_x(AH1(a)) + #define AH4_(a) AH4_x(AH1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AW1_x(AW1 a){return AW1(a);} + AW2 AW2_x(AW1 a){return AW2(a,a);} + AW3 AW3_x(AW1 a){return AW3(a,a,a);} + AW4 AW4_x(AW1 a){return AW4(a,a,a,a);} + #define AW1_(a) AW1_x(AW1(a)) + #define AW2_(a) AW2_x(AW1(a)) + #define AW3_(a) AW3_x(AW1(a)) + #define AW4_(a) AW4_x(AW1(a)) +//============================================================================================================================== + AW1 AAbsSW1(AW1 a){return AW1(abs(ASW1(a)));} + AW2 AAbsSW2(AW2 a){return AW2(abs(ASW2(a)));} + AW3 AAbsSW3(AW3 a){return AW3(abs(ASW3(a)));} + AW4 AAbsSW4(AW4 a){return AW4(abs(ASW4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AClampH1(AH1 x,AH1 n,AH1 m){return clamp(x,n,m);} + AH2 AClampH2(AH2 x,AH2 n,AH2 m){return clamp(x,n,m);} + AH3 AClampH3(AH3 x,AH3 n,AH3 m){return clamp(x,n,m);} + AH4 AClampH4(AH4 x,AH4 n,AH4 m){return clamp(x,n,m);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFractH1(AH1 x){return fract(x);} + AH2 AFractH2(AH2 x){return fract(x);} + AH3 AFractH3(AH3 x){return fract(x);} + AH4 AFractH4(AH4 x){return fract(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ALerpH1(AH1 x,AH1 y,AH1 a){return mix(x,y,a);} + AH2 ALerpH2(AH2 x,AH2 y,AH2 a){return mix(x,y,a);} + AH3 ALerpH3(AH3 x,AH3 y,AH3 a){return mix(x,y,a);} + AH4 ALerpH4(AH4 x,AH4 y,AH4 a){return mix(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + // No packed version of max3. + AH1 AMax3H1(AH1 x,AH1 y,AH1 z){return max(x,max(y,z));} + AH2 AMax3H2(AH2 x,AH2 y,AH2 z){return max(x,max(y,z));} + AH3 AMax3H3(AH3 x,AH3 y,AH3 z){return max(x,max(y,z));} + AH4 AMax3H4(AH4 x,AH4 y,AH4 z){return max(x,max(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMaxSW1(AW1 a,AW1 b){return AW1(max(ASU1(a),ASU1(b)));} + AW2 AMaxSW2(AW2 a,AW2 b){return AW2(max(ASU2(a),ASU2(b)));} + AW3 AMaxSW3(AW3 a,AW3 b){return AW3(max(ASU3(a),ASU3(b)));} + AW4 AMaxSW4(AW4 a,AW4 b){return AW4(max(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + // No packed version of min3. + AH1 AMin3H1(AH1 x,AH1 y,AH1 z){return min(x,min(y,z));} + AH2 AMin3H2(AH2 x,AH2 y,AH2 z){return min(x,min(y,z));} + AH3 AMin3H3(AH3 x,AH3 y,AH3 z){return min(x,min(y,z));} + AH4 AMin3H4(AH4 x,AH4 y,AH4 z){return min(x,min(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMinSW1(AW1 a,AW1 b){return AW1(min(ASU1(a),ASU1(b)));} + AW2 AMinSW2(AW2 a,AW2 b){return AW2(min(ASU2(a),ASU2(b)));} + AW3 AMinSW3(AW3 a,AW3 b){return AW3(min(ASU3(a),ASU3(b)));} + AW4 AMinSW4(AW4 a,AW4 b){return AW4(min(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARcpH1(AH1 x){return AH1_(1.0)/x;} + AH2 ARcpH2(AH2 x){return AH2_(1.0)/x;} + AH3 ARcpH3(AH3 x){return AH3_(1.0)/x;} + AH4 ARcpH4(AH4 x){return AH4_(1.0)/x;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARsqH1(AH1 x){return AH1_(1.0)/sqrt(x);} + AH2 ARsqH2(AH2 x){return AH2_(1.0)/sqrt(x);} + AH3 ARsqH3(AH3 x){return AH3_(1.0)/sqrt(x);} + AH4 ARsqH4(AH4 x){return AH4_(1.0)/sqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASatH1(AH1 x){return clamp(x,AH1_(0.0),AH1_(1.0));} + AH2 ASatH2(AH2 x){return clamp(x,AH2_(0.0),AH2_(1.0));} + AH3 ASatH3(AH3 x){return clamp(x,AH3_(0.0),AH3_(1.0));} + AH4 ASatH4(AH4 x){return clamp(x,AH4_(0.0),AH4_(1.0));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AShrSW1(AW1 a,AW1 b){return AW1(ASW1(a)>>ASW1(b));} + AW2 AShrSW2(AW2 a,AW2 b){return AW2(ASW2(a)>>ASW2(b));} + AW3 AShrSW3(AW3 a,AW3 b){return AW3(ASW3(a)>>ASW3(b));} + AW4 AShrSW4(AW4 a,AW4 b){return AW4(ASW4(a)>>ASW4(b));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL DOUBLE +//============================================================================================================================== + #ifdef A_DUBL + #define AD1 double + #define AD2 dvec2 + #define AD3 dvec3 + #define AD4 dvec4 +//------------------------------------------------------------------------------------------------------------------------------ + AD1 AD1_x(AD1 a){return AD1(a);} + AD2 AD2_x(AD1 a){return AD2(a,a);} + AD3 AD3_x(AD1 a){return AD3(a,a,a);} + AD4 AD4_x(AD1 a){return AD4(a,a,a,a);} + #define AD1_(a) AD1_x(AD1(a)) + #define AD2_(a) AD2_x(AD1(a)) + #define AD3_(a) AD3_x(AD1(a)) + #define AD4_(a) AD4_x(AD1(a)) +//============================================================================================================================== + AD1 AFractD1(AD1 x){return fract(x);} + AD2 AFractD2(AD2 x){return fract(x);} + AD3 AFractD3(AD3 x){return fract(x);} + AD4 AFractD4(AD4 x){return fract(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ALerpD1(AD1 x,AD1 y,AD1 a){return mix(x,y,a);} + AD2 ALerpD2(AD2 x,AD2 y,AD2 a){return mix(x,y,a);} + AD3 ALerpD3(AD3 x,AD3 y,AD3 a){return mix(x,y,a);} + AD4 ALerpD4(AD4 x,AD4 y,AD4 a){return mix(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARcpD1(AD1 x){return AD1_(1.0)/x;} + AD2 ARcpD2(AD2 x){return AD2_(1.0)/x;} + AD3 ARcpD3(AD3 x){return AD3_(1.0)/x;} + AD4 ARcpD4(AD4 x){return AD4_(1.0)/x;} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARsqD1(AD1 x){return AD1_(1.0)/sqrt(x);} + AD2 ARsqD2(AD2 x){return AD2_(1.0)/sqrt(x);} + AD3 ARsqD3(AD3 x){return AD3_(1.0)/sqrt(x);} + AD4 ARsqD4(AD4 x){return AD4_(1.0)/sqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ASatD1(AD1 x){return clamp(x,AD1_(0.0),AD1_(1.0));} + AD2 ASatD2(AD2 x){return clamp(x,AD2_(0.0),AD2_(1.0));} + AD3 ASatD3(AD3 x){return clamp(x,AD3_(0.0),AD3_(1.0));} + AD4 ASatD4(AD4 x){return clamp(x,AD4_(0.0),AD4_(1.0));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// GLSL LONG +//============================================================================================================================== + #ifdef A_LONG + #define AL1 uint64_t + #define AL2 u64vec2 + #define AL3 u64vec3 + #define AL4 u64vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASL1 int64_t + #define ASL2 i64vec2 + #define ASL3 i64vec3 + #define ASL4 i64vec4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AL1_AU2(x) packUint2x32(AU2(x)) + #define AU2_AL1(x) unpackUint2x32(AL1(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AL1 AL1_x(AL1 a){return AL1(a);} + AL2 AL2_x(AL1 a){return AL2(a,a);} + AL3 AL3_x(AL1 a){return AL3(a,a,a);} + AL4 AL4_x(AL1 a){return AL4(a,a,a,a);} + #define AL1_(a) AL1_x(AL1(a)) + #define AL2_(a) AL2_x(AL1(a)) + #define AL3_(a) AL3_x(AL1(a)) + #define AL4_(a) AL4_x(AL1(a)) +//============================================================================================================================== + AL1 AAbsSL1(AL1 a){return AL1(abs(ASL1(a)));} + AL2 AAbsSL2(AL2 a){return AL2(abs(ASL2(a)));} + AL3 AAbsSL3(AL3 a){return AL3(abs(ASL3(a)));} + AL4 AAbsSL4(AL4 a){return AL4(abs(ASL4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AL1 AMaxSL1(AL1 a,AL1 b){return AL1(max(ASU1(a),ASU1(b)));} + AL2 AMaxSL2(AL2 a,AL2 b){return AL2(max(ASU2(a),ASU2(b)));} + AL3 AMaxSL3(AL3 a,AL3 b){return AL3(max(ASU3(a),ASU3(b)));} + AL4 AMaxSL4(AL4 a,AL4 b){return AL4(max(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AL1 AMinSL1(AL1 a,AL1 b){return AL1(min(ASU1(a),ASU1(b)));} + AL2 AMinSL2(AL2 a,AL2 b){return AL2(min(ASU2(a),ASU2(b)));} + AL3 AMinSL3(AL3 a,AL3 b){return AL3(min(ASU3(a),ASU3(b)));} + AL4 AMinSL4(AL4 a,AL4 b){return AL4(min(ASU4(a),ASU4(b)));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// WAVE OPERATIONS +//============================================================================================================================== + #ifdef A_WAVE + // Where 'x' must be a compile time literal. + AF1 AWaveXorF1(AF1 v,AU1 x){return subgroupShuffleXor(v,x);} + AF2 AWaveXorF2(AF2 v,AU1 x){return subgroupShuffleXor(v,x);} + AF3 AWaveXorF3(AF3 v,AU1 x){return subgroupShuffleXor(v,x);} + AF4 AWaveXorF4(AF4 v,AU1 x){return subgroupShuffleXor(v,x);} + AU1 AWaveXorU1(AU1 v,AU1 x){return subgroupShuffleXor(v,x);} + AU2 AWaveXorU2(AU2 v,AU1 x){return subgroupShuffleXor(v,x);} + AU3 AWaveXorU3(AU3 v,AU1 x){return subgroupShuffleXor(v,x);} + AU4 AWaveXorU4(AU4 v,AU1 x){return subgroupShuffleXor(v,x);} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + AH2 AWaveXorH2(AH2 v,AU1 x){return AH2_AU1(subgroupShuffleXor(AU1_AH2(v),x));} + AH4 AWaveXorH4(AH4 v,AU1 x){return AH4_AU2(subgroupShuffleXor(AU2_AH4(v),x));} + AW2 AWaveXorW2(AW2 v,AU1 x){return AW2_AU1(subgroupShuffleXor(AU1_AW2(v),x));} + AW4 AWaveXorW4(AW4 v,AU1 x){return AW4_AU2(subgroupShuffleXor(AU2_AW4(v),x));} + #endif + #endif +//============================================================================================================================== +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// HLSL +// +// +//============================================================================================================================== +#if defined(A_HLSL) && defined(A_GPU) + #ifdef A_HLSL_6_2 + #define AP1 bool + #define AP2 bool2 + #define AP3 bool3 + #define AP4 bool4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF1 float32_t + #define AF2 float32_t2 + #define AF3 float32_t3 + #define AF4 float32_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1 uint32_t + #define AU2 uint32_t2 + #define AU3 uint32_t3 + #define AU4 uint32_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASU1 int32_t + #define ASU2 int32_t2 + #define ASU3 int32_t3 + #define ASU4 int32_t4 + #else + #define AP1 bool + #define AP2 bool2 + #define AP3 bool3 + #define AP4 bool4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AF1 float + #define AF2 float2 + #define AF3 float3 + #define AF4 float4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1 uint + #define AU2 uint2 + #define AU3 uint3 + #define AU4 uint4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASU1 int + #define ASU2 int2 + #define ASU3 int3 + #define ASU4 int4 + #endif +//============================================================================================================================== + #define AF1_AU1(x) asfloat(AU1(x)) + #define AF2_AU2(x) asfloat(AU2(x)) + #define AF3_AU3(x) asfloat(AU3(x)) + #define AF4_AU4(x) asfloat(AU4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AU1_AF1(x) asuint(AF1(x)) + #define AU2_AF2(x) asuint(AF2(x)) + #define AU3_AF3(x) asuint(AF3(x)) + #define AU4_AF4(x) asuint(AF4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH1_AF1_x(AF1 a){return f32tof16(a);} + #define AU1_AH1_AF1(a) AU1_AH1_AF1_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH2_AF2_x(AF2 a){return f32tof16(a.x)|(f32tof16(a.y)<<16);} + #define AU1_AH2_AF2(a) AU1_AH2_AF2_x(AF2(a)) + #define AU1_AB4Unorm_AF4(x) D3DCOLORtoUBYTE4(AF4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AF2 AF2_AH2_AU1_x(AU1 x){return AF2(f16tof32(x&0xFFFF),f16tof32(x>>16));} + #define AF2_AH2_AU1(x) AF2_AH2_AU1_x(AU1(x)) +//============================================================================================================================== + AF1 AF1_x(AF1 a){return AF1(a);} + AF2 AF2_x(AF1 a){return AF2(a,a);} + AF3 AF3_x(AF1 a){return AF3(a,a,a);} + AF4 AF4_x(AF1 a){return AF4(a,a,a,a);} + #define AF1_(a) AF1_x(AF1(a)) + #define AF2_(a) AF2_x(AF1(a)) + #define AF3_(a) AF3_x(AF1(a)) + #define AF4_(a) AF4_x(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_x(AU1 a){return AU1(a);} + AU2 AU2_x(AU1 a){return AU2(a,a);} + AU3 AU3_x(AU1 a){return AU3(a,a,a);} + AU4 AU4_x(AU1 a){return AU4(a,a,a,a);} + #define AU1_(a) AU1_x(AU1(a)) + #define AU2_(a) AU2_x(AU1(a)) + #define AU3_(a) AU3_x(AU1(a)) + #define AU4_(a) AU4_x(AU1(a)) +//============================================================================================================================== + AU1 AAbsSU1(AU1 a){return AU1(abs(ASU1(a)));} + AU2 AAbsSU2(AU2 a){return AU2(abs(ASU2(a)));} + AU3 AAbsSU3(AU3 a){return AU3(abs(ASU3(a)));} + AU4 AAbsSU4(AU4 a){return AU4(abs(ASU4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 ABfe(AU1 src,AU1 off,AU1 bits){AU1 mask=(1u<>off)&mask;} + AU1 ABfi(AU1 src,AU1 ins,AU1 mask){return (ins&mask)|(src&(~mask));} + AU1 ABfiM(AU1 src,AU1 ins,AU1 bits){AU1 mask=(1u<>ASU1(b));} + AU2 AShrSU2(AU2 a,AU2 b){return AU2(ASU2(a)>>ASU2(b));} + AU3 AShrSU3(AU3 a,AU3 b){return AU3(ASU3(a)>>ASU3(b));} + AU4 AShrSU4(AU4 a,AU4 b){return AU4(ASU4(a)>>ASU4(b));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HLSL BYTE +//============================================================================================================================== + #ifdef A_BYTE + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HLSL HALF +//============================================================================================================================== + #ifdef A_HALF + #ifdef A_HLSL_6_2 + #define AH1 float16_t + #define AH2 float16_t2 + #define AH3 float16_t3 + #define AH4 float16_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AW1 uint16_t + #define AW2 uint16_t2 + #define AW3 uint16_t3 + #define AW4 uint16_t4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASW1 int16_t + #define ASW2 int16_t2 + #define ASW3 int16_t3 + #define ASW4 int16_t4 + #else + #define AH1 min16float + #define AH2 min16float2 + #define AH3 min16float3 + #define AH4 min16float4 +//------------------------------------------------------------------------------------------------------------------------------ + #define AW1 min16uint + #define AW2 min16uint2 + #define AW3 min16uint3 + #define AW4 min16uint4 +//------------------------------------------------------------------------------------------------------------------------------ + #define ASW1 min16int + #define ASW2 min16int2 + #define ASW3 min16int3 + #define ASW4 min16int4 + #endif +//============================================================================================================================== + // Need to use manual unpack to get optimal execution (don't use packed types in buffers directly). + // Unpack requires this pattern: https://gpuopen.com/first-steps-implementing-fp16/ + AH2 AH2_AU1_x(AU1 x){AF2 t=f16tof32(AU2(x&0xFFFF,x>>16));return AH2(t);} + AH4 AH4_AU2_x(AU2 x){return AH4(AH2_AU1_x(x.x),AH2_AU1_x(x.y));} + AW2 AW2_AU1_x(AU1 x){AU2 t=AU2(x&0xFFFF,x>>16);return AW2(t);} + AW4 AW4_AU2_x(AU2 x){return AW4(AW2_AU1_x(x.x),AW2_AU1_x(x.y));} + #define AH2_AU1(x) AH2_AU1_x(AU1(x)) + #define AH4_AU2(x) AH4_AU2_x(AU2(x)) + #define AW2_AU1(x) AW2_AU1_x(AU1(x)) + #define AW4_AU2(x) AW4_AU2_x(AU2(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH2_x(AH2 x){return f32tof16(x.x)+(f32tof16(x.y)<<16);} + AU2 AU2_AH4_x(AH4 x){return AU2(AU1_AH2_x(x.xy),AU1_AH2_x(x.zw));} + AU1 AU1_AW2_x(AW2 x){return AU1(x.x)+(AU1(x.y)<<16);} + AU2 AU2_AW4_x(AW4 x){return AU2(AU1_AW2_x(x.xy),AU1_AW2_x(x.zw));} + #define AU1_AH2(x) AU1_AH2_x(AH2(x)) + #define AU2_AH4(x) AU2_AH4_x(AH4(x)) + #define AU1_AW2(x) AU1_AW2_x(AW2(x)) + #define AU2_AW4(x) AU2_AW4_x(AW4(x)) +//============================================================================================================================== + #if defined(A_HLSL_6_2) && !defined(A_NO_16_BIT_CAST) + #define AW1_AH1(x) asuint16(x) + #define AW2_AH2(x) asuint16(x) + #define AW3_AH3(x) asuint16(x) + #define AW4_AH4(x) asuint16(x) + #else + #define AW1_AH1(a) AW1(f32tof16(AF1(a))) + #define AW2_AH2(a) AW2(AW1_AH1((a).x),AW1_AH1((a).y)) + #define AW3_AH3(a) AW3(AW1_AH1((a).x),AW1_AH1((a).y),AW1_AH1((a).z)) + #define AW4_AH4(a) AW4(AW1_AH1((a).x),AW1_AH1((a).y),AW1_AH1((a).z),AW1_AH1((a).w)) + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #if defined(A_HLSL_6_2) && !defined(A_NO_16_BIT_CAST) + #define AH1_AW1(x) asfloat16(x) + #define AH2_AW2(x) asfloat16(x) + #define AH3_AW3(x) asfloat16(x) + #define AH4_AW4(x) asfloat16(x) + #else + #define AH1_AW1(a) AH1(f16tof32(AU1(a))) + #define AH2_AW2(a) AH2(AH1_AW1((a).x),AH1_AW1((a).y)) + #define AH3_AW3(a) AH3(AH1_AW1((a).x),AH1_AW1((a).y),AH1_AW1((a).z)) + #define AH4_AW4(a) AH4(AH1_AW1((a).x),AH1_AW1((a).y),AH1_AW1((a).z),AH1_AW1((a).w)) + #endif +//============================================================================================================================== + AH1 AH1_x(AH1 a){return AH1(a);} + AH2 AH2_x(AH1 a){return AH2(a,a);} + AH3 AH3_x(AH1 a){return AH3(a,a,a);} + AH4 AH4_x(AH1 a){return AH4(a,a,a,a);} + #define AH1_(a) AH1_x(AH1(a)) + #define AH2_(a) AH2_x(AH1(a)) + #define AH3_(a) AH3_x(AH1(a)) + #define AH4_(a) AH4_x(AH1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AW1_x(AW1 a){return AW1(a);} + AW2 AW2_x(AW1 a){return AW2(a,a);} + AW3 AW3_x(AW1 a){return AW3(a,a,a);} + AW4 AW4_x(AW1 a){return AW4(a,a,a,a);} + #define AW1_(a) AW1_x(AW1(a)) + #define AW2_(a) AW2_x(AW1(a)) + #define AW3_(a) AW3_x(AW1(a)) + #define AW4_(a) AW4_x(AW1(a)) +//============================================================================================================================== + AW1 AAbsSW1(AW1 a){return AW1(abs(ASW1(a)));} + AW2 AAbsSW2(AW2 a){return AW2(abs(ASW2(a)));} + AW3 AAbsSW3(AW3 a){return AW3(abs(ASW3(a)));} + AW4 AAbsSW4(AW4 a){return AW4(abs(ASW4(a)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AClampH1(AH1 x,AH1 n,AH1 m){return max(n,min(x,m));} + AH2 AClampH2(AH2 x,AH2 n,AH2 m){return max(n,min(x,m));} + AH3 AClampH3(AH3 x,AH3 n,AH3 m){return max(n,min(x,m));} + AH4 AClampH4(AH4 x,AH4 n,AH4 m){return max(n,min(x,m));} +//------------------------------------------------------------------------------------------------------------------------------ + // V_FRACT_F16 (note DX frac() is different). + AH1 AFractH1(AH1 x){return x-floor(x);} + AH2 AFractH2(AH2 x){return x-floor(x);} + AH3 AFractH3(AH3 x){return x-floor(x);} + AH4 AFractH4(AH4 x){return x-floor(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ALerpH1(AH1 x,AH1 y,AH1 a){return lerp(x,y,a);} + AH2 ALerpH2(AH2 x,AH2 y,AH2 a){return lerp(x,y,a);} + AH3 ALerpH3(AH3 x,AH3 y,AH3 a){return lerp(x,y,a);} + AH4 ALerpH4(AH4 x,AH4 y,AH4 a){return lerp(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AMax3H1(AH1 x,AH1 y,AH1 z){return max(x,max(y,z));} + AH2 AMax3H2(AH2 x,AH2 y,AH2 z){return max(x,max(y,z));} + AH3 AMax3H3(AH3 x,AH3 y,AH3 z){return max(x,max(y,z));} + AH4 AMax3H4(AH4 x,AH4 y,AH4 z){return max(x,max(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMaxSW1(AW1 a,AW1 b){return AW1(max(ASU1(a),ASU1(b)));} + AW2 AMaxSW2(AW2 a,AW2 b){return AW2(max(ASU2(a),ASU2(b)));} + AW3 AMaxSW3(AW3 a,AW3 b){return AW3(max(ASU3(a),ASU3(b)));} + AW4 AMaxSW4(AW4 a,AW4 b){return AW4(max(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AMin3H1(AH1 x,AH1 y,AH1 z){return min(x,min(y,z));} + AH2 AMin3H2(AH2 x,AH2 y,AH2 z){return min(x,min(y,z));} + AH3 AMin3H3(AH3 x,AH3 y,AH3 z){return min(x,min(y,z));} + AH4 AMin3H4(AH4 x,AH4 y,AH4 z){return min(x,min(y,z));} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AMinSW1(AW1 a,AW1 b){return AW1(min(ASU1(a),ASU1(b)));} + AW2 AMinSW2(AW2 a,AW2 b){return AW2(min(ASU2(a),ASU2(b)));} + AW3 AMinSW3(AW3 a,AW3 b){return AW3(min(ASU3(a),ASU3(b)));} + AW4 AMinSW4(AW4 a,AW4 b){return AW4(min(ASU4(a),ASU4(b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARcpH1(AH1 x){return rcp(x);} + AH2 ARcpH2(AH2 x){return rcp(x);} + AH3 ARcpH3(AH3 x){return rcp(x);} + AH4 ARcpH4(AH4 x){return rcp(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ARsqH1(AH1 x){return rsqrt(x);} + AH2 ARsqH2(AH2 x){return rsqrt(x);} + AH3 ARsqH3(AH3 x){return rsqrt(x);} + AH4 ARsqH4(AH4 x){return rsqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASatH1(AH1 x){return saturate(x);} + AH2 ASatH2(AH2 x){return saturate(x);} + AH3 ASatH3(AH3 x){return saturate(x);} + AH4 ASatH4(AH4 x){return saturate(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AShrSW1(AW1 a,AW1 b){return AW1(ASW1(a)>>ASW1(b));} + AW2 AShrSW2(AW2 a,AW2 b){return AW2(ASW2(a)>>ASW2(b));} + AW3 AShrSW3(AW3 a,AW3 b){return AW3(ASW3(a)>>ASW3(b));} + AW4 AShrSW4(AW4 a,AW4 b){return AW4(ASW4(a)>>ASW4(b));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HLSL DOUBLE +//============================================================================================================================== + #ifdef A_DUBL + #ifdef A_HLSL_6_2 + #define AD1 float64_t + #define AD2 float64_t2 + #define AD3 float64_t3 + #define AD4 float64_t4 + #else + #define AD1 double + #define AD2 double2 + #define AD3 double3 + #define AD4 double4 + #endif +//------------------------------------------------------------------------------------------------------------------------------ + AD1 AD1_x(AD1 a){return AD1(a);} + AD2 AD2_x(AD1 a){return AD2(a,a);} + AD3 AD3_x(AD1 a){return AD3(a,a,a);} + AD4 AD4_x(AD1 a){return AD4(a,a,a,a);} + #define AD1_(a) AD1_x(AD1(a)) + #define AD2_(a) AD2_x(AD1(a)) + #define AD3_(a) AD3_x(AD1(a)) + #define AD4_(a) AD4_x(AD1(a)) +//============================================================================================================================== + AD1 AFractD1(AD1 a){return a-floor(a);} + AD2 AFractD2(AD2 a){return a-floor(a);} + AD3 AFractD3(AD3 a){return a-floor(a);} + AD4 AFractD4(AD4 a){return a-floor(a);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ALerpD1(AD1 x,AD1 y,AD1 a){return lerp(x,y,a);} + AD2 ALerpD2(AD2 x,AD2 y,AD2 a){return lerp(x,y,a);} + AD3 ALerpD3(AD3 x,AD3 y,AD3 a){return lerp(x,y,a);} + AD4 ALerpD4(AD4 x,AD4 y,AD4 a){return lerp(x,y,a);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARcpD1(AD1 x){return rcp(x);} + AD2 ARcpD2(AD2 x){return rcp(x);} + AD3 ARcpD3(AD3 x){return rcp(x);} + AD4 ARcpD4(AD4 x){return rcp(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ARsqD1(AD1 x){return rsqrt(x);} + AD2 ARsqD2(AD2 x){return rsqrt(x);} + AD3 ARsqD3(AD3 x){return rsqrt(x);} + AD4 ARsqD4(AD4 x){return rsqrt(x);} +//------------------------------------------------------------------------------------------------------------------------------ + AD1 ASatD1(AD1 x){return saturate(x);} + AD2 ASatD2(AD2 x){return saturate(x);} + AD3 ASatD3(AD3 x){return saturate(x);} + AD4 ASatD4(AD4 x){return saturate(x);} + #endif +//============================================================================================================================== +// HLSL WAVE +//============================================================================================================================== + #ifdef A_WAVE + // Where 'x' must be a compile time literal. + AF1 AWaveXorF1(AF1 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AF2 AWaveXorF2(AF2 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AF3 AWaveXorF3(AF3 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AF4 AWaveXorF4(AF4 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU1 AWaveXorU1(AU1 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU2 AWaveXorU1(AU2 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU3 AWaveXorU1(AU3 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} + AU4 AWaveXorU1(AU4 v,AU1 x){return WaveReadLaneAt(v,WaveGetLaneIndex()^x);} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + AH2 AWaveXorH2(AH2 v,AU1 x){return AH2_AU1(WaveReadLaneAt(AU1_AH2(v),WaveGetLaneIndex()^x));} + AH4 AWaveXorH4(AH4 v,AU1 x){return AH4_AU2(WaveReadLaneAt(AU2_AH4(v),WaveGetLaneIndex()^x));} + AW2 AWaveXorW2(AW2 v,AU1 x){return AW2_AU1(WaveReadLaneAt(AU1_AW2(v),WaveGetLaneIndex()^x));} + AW4 AWaveXorW4(AW4 v,AU1 x){return AW4_AU1(WaveReadLaneAt(AU1_AW4(v),WaveGetLaneIndex()^x));} + #endif + #endif +//============================================================================================================================== +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// GPU COMMON +// +// +//============================================================================================================================== +#ifdef A_GPU + // Negative and positive infinity. + #define A_INFP_F AF1_AU1(0x7f800000u) + #define A_INFN_F AF1_AU1(0xff800000u) +//------------------------------------------------------------------------------------------------------------------------------ + // Copy sign from 's' to positive 'd'. + AF1 ACpySgnF1(AF1 d,AF1 s){return AF1_AU1(AU1_AF1(d)|(AU1_AF1(s)&AU1_(0x80000000u)));} + AF2 ACpySgnF2(AF2 d,AF2 s){return AF2_AU2(AU2_AF2(d)|(AU2_AF2(s)&AU2_(0x80000000u)));} + AF3 ACpySgnF3(AF3 d,AF3 s){return AF3_AU3(AU3_AF3(d)|(AU3_AF3(s)&AU3_(0x80000000u)));} + AF4 ACpySgnF4(AF4 d,AF4 s){return AF4_AU4(AU4_AF4(d)|(AU4_AF4(s)&AU4_(0x80000000u)));} +//------------------------------------------------------------------------------------------------------------------------------ + // Single operation to return (useful to create a mask to use in lerp for branch free logic), + // m=NaN := 0 + // m>=0 := 0 + // m<0 := 1 + // Uses the following useful floating point logic, + // saturate(+a*(-INF)==-INF) := 0 + // saturate( 0*(-INF)== NaN) := 0 + // saturate(-a*(-INF)==+INF) := 1 + AF1 ASignedF1(AF1 m){return ASatF1(m*AF1_(A_INFN_F));} + AF2 ASignedF2(AF2 m){return ASatF2(m*AF2_(A_INFN_F));} + AF3 ASignedF3(AF3 m){return ASatF3(m*AF3_(A_INFN_F));} + AF4 ASignedF4(AF4 m){return ASatF4(m*AF4_(A_INFN_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AGtZeroF1(AF1 m){return ASatF1(m*AF1_(A_INFP_F));} + AF2 AGtZeroF2(AF2 m){return ASatF2(m*AF2_(A_INFP_F));} + AF3 AGtZeroF3(AF3 m){return ASatF3(m*AF3_(A_INFP_F));} + AF4 AGtZeroF4(AF4 m){return ASatF4(m*AF4_(A_INFP_F));} +//============================================================================================================================== + #ifdef A_HALF + #ifdef A_HLSL_6_2 + #define A_INFP_H AH1_AW1((uint16_t)0x7c00u) + #define A_INFN_H AH1_AW1((uint16_t)0xfc00u) + #else + #define A_INFP_H AH1_AW1(0x7c00u) + #define A_INFN_H AH1_AW1(0xfc00u) + #endif + +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ACpySgnH1(AH1 d,AH1 s){return AH1_AW1(AW1_AH1(d)|(AW1_AH1(s)&AW1_(0x8000u)));} + AH2 ACpySgnH2(AH2 d,AH2 s){return AH2_AW2(AW2_AH2(d)|(AW2_AH2(s)&AW2_(0x8000u)));} + AH3 ACpySgnH3(AH3 d,AH3 s){return AH3_AW3(AW3_AH3(d)|(AW3_AH3(s)&AW3_(0x8000u)));} + AH4 ACpySgnH4(AH4 d,AH4 s){return AH4_AW4(AW4_AH4(d)|(AW4_AH4(s)&AW4_(0x8000u)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASignedH1(AH1 m){return ASatH1(m*AH1_(A_INFN_H));} + AH2 ASignedH2(AH2 m){return ASatH2(m*AH2_(A_INFN_H));} + AH3 ASignedH3(AH3 m){return ASatH3(m*AH3_(A_INFN_H));} + AH4 ASignedH4(AH4 m){return ASatH4(m*AH4_(A_INFN_H));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AGtZeroH1(AH1 m){return ASatH1(m*AH1_(A_INFP_H));} + AH2 AGtZeroH2(AH2 m){return ASatH2(m*AH2_(A_INFP_H));} + AH3 AGtZeroH3(AH3 m){return ASatH3(m*AH3_(A_INFP_H));} + AH4 AGtZeroH4(AH4 m){return ASatH4(m*AH4_(A_INFP_H));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [FIS] FLOAT INTEGER SORTABLE +//------------------------------------------------------------------------------------------------------------------------------ +// Float to integer sortable. +// - If sign bit=0, flip the sign bit (positives). +// - If sign bit=1, flip all bits (negatives). +// Integer sortable to float. +// - If sign bit=1, flip the sign bit (positives). +// - If sign bit=0, flip all bits (negatives). +// Has nice side effects. +// - Larger integers are more positive values. +// - Float zero is mapped to center of integers (so clear to integer zero is a nice default for atomic max usage). +// Burns 3 ops for conversion {shift,or,xor}. +//============================================================================================================================== + AU1 AFisToU1(AU1 x){return x^(( AShrSU1(x,AU1_(31)))|AU1_(0x80000000));} + AU1 AFisFromU1(AU1 x){return x^((~AShrSU1(x,AU1_(31)))|AU1_(0x80000000));} +//------------------------------------------------------------------------------------------------------------------------------ + // Just adjust high 16-bit value (useful when upper part of 32-bit word is a 16-bit float value). + AU1 AFisToHiU1(AU1 x){return x^(( AShrSU1(x,AU1_(15)))|AU1_(0x80000000));} + AU1 AFisFromHiU1(AU1 x){return x^((~AShrSU1(x,AU1_(15)))|AU1_(0x80000000));} +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + AW1 AFisToW1(AW1 x){return x^(( AShrSW1(x,AW1_(15)))|AW1_(0x8000));} + AW1 AFisFromW1(AW1 x){return x^((~AShrSW1(x,AW1_(15)))|AW1_(0x8000));} +//------------------------------------------------------------------------------------------------------------------------------ + AW2 AFisToW2(AW2 x){return x^(( AShrSW2(x,AW2_(15)))|AW2_(0x8000));} + AW2 AFisFromW2(AW2 x){return x^((~AShrSW2(x,AW2_(15)))|AW2_(0x8000));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [PERM] V_PERM_B32 +//------------------------------------------------------------------------------------------------------------------------------ +// Support for V_PERM_B32 started in the 3rd generation of GCN. +//------------------------------------------------------------------------------------------------------------------------------ +// yyyyxxxx - The 'i' input. +// 76543210 +// ======== +// HGFEDCBA - Naming on permutation. +//------------------------------------------------------------------------------------------------------------------------------ +// TODO +// ==== +// - Make sure compiler optimizes this. +//============================================================================================================================== + #ifdef A_HALF + AU1 APerm0E0A(AU2 i){return((i.x )&0xffu)|((i.y<<16)&0xff0000u);} + AU1 APerm0F0B(AU2 i){return((i.x>> 8)&0xffu)|((i.y<< 8)&0xff0000u);} + AU1 APerm0G0C(AU2 i){return((i.x>>16)&0xffu)|((i.y )&0xff0000u);} + AU1 APerm0H0D(AU2 i){return((i.x>>24)&0xffu)|((i.y>> 8)&0xff0000u);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 APermHGFA(AU2 i){return((i.x )&0x000000ffu)|(i.y&0xffffff00u);} + AU1 APermHGFC(AU2 i){return((i.x>>16)&0x000000ffu)|(i.y&0xffffff00u);} + AU1 APermHGAE(AU2 i){return((i.x<< 8)&0x0000ff00u)|(i.y&0xffff00ffu);} + AU1 APermHGCE(AU2 i){return((i.x>> 8)&0x0000ff00u)|(i.y&0xffff00ffu);} + AU1 APermHAFE(AU2 i){return((i.x<<16)&0x00ff0000u)|(i.y&0xff00ffffu);} + AU1 APermHCFE(AU2 i){return((i.x )&0x00ff0000u)|(i.y&0xff00ffffu);} + AU1 APermAGFE(AU2 i){return((i.x<<24)&0xff000000u)|(i.y&0x00ffffffu);} + AU1 APermCGFE(AU2 i){return((i.x<< 8)&0xff000000u)|(i.y&0x00ffffffu);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 APermGCEA(AU2 i){return((i.x)&0x00ff00ffu)|((i.y<<8)&0xff00ff00u);} + AU1 APermGECA(AU2 i){return(((i.x)&0xffu)|((i.x>>8)&0xff00u)|((i.y<<16)&0xff0000u)|((i.y<<8)&0xff000000u));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [BUC] BYTE UNSIGNED CONVERSION +//------------------------------------------------------------------------------------------------------------------------------ +// Designed to use the optimal conversion, enables the scaling to possibly be factored into other computation. +// Works on a range of {0 to A_BUC_<32,16>}, for <32-bit, and 16-bit> respectively. +//------------------------------------------------------------------------------------------------------------------------------ +// OPCODE NOTES +// ============ +// GCN does not do UNORM or SNORM for bytes in opcodes. +// - V_CVT_F32_UBYTE{0,1,2,3} - Unsigned byte to float. +// - V_CVT_PKACC_U8_F32 - Float to unsigned byte (does bit-field insert into 32-bit integer). +// V_PERM_B32 does byte packing with ability to zero fill bytes as well. +// - Can pull out byte values from two sources, and zero fill upper 8-bits of packed hi and lo. +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABuc{0,1,2,3}{To,From}U1() - Designed for V_CVT_F32_UBYTE* and V_CVT_PKACCUM_U8_F32 ops. +// ==== ===== +// 0 : 0 +// 1 : 1 +// ... +// 255 : 255 +// : 256 (just outside the encoding range) +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABuc{0,1,2,3}{To,From}U2() - Designed for 16-bit denormal tricks and V_PERM_B32. +// ==== ===== +// 0 : 0 +// 1 : 1/512 +// 2 : 1/256 +// ... +// 64 : 1/8 +// 128 : 1/4 +// 255 : 255/512 +// : 1/2 (just outside the encoding range) +//------------------------------------------------------------------------------------------------------------------------------ +// OPTIMAL IMPLEMENTATIONS ON AMD ARCHITECTURES +// ============================================ +// r=ABuc0FromU1(i) +// V_CVT_F32_UBYTE0 r,i +// -------------------------------------------- +// r=ABuc0ToU1(d,i) +// V_CVT_PKACCUM_U8_F32 r,i,0,d +// -------------------------------------------- +// d=ABuc0FromU2(i) +// Where 'k0' is an SGPR with 0x0E0A +// Where 'k1' is an SGPR with {32768.0} packed into the lower 16-bits +// V_PERM_B32 d,i.x,i.y,k0 +// V_PK_FMA_F16 d,d,k1.x,0 +// -------------------------------------------- +// r=ABuc0ToU2(d,i) +// Where 'k0' is an SGPR with {1.0/32768.0} packed into the lower 16-bits +// Where 'k1' is an SGPR with 0x???? +// Where 'k2' is an SGPR with 0x???? +// V_PK_FMA_F16 i,i,k0.x,0 +// V_PERM_B32 r.x,i,i,k1 +// V_PERM_B32 r.y,i,i,k2 +//============================================================================================================================== + // Peak range for 32-bit and 16-bit operations. + #define A_BUC_32 (255.0) + #define A_BUC_16 (255.0/512.0) +//============================================================================================================================== + #if 1 + // Designed to be one V_CVT_PKACCUM_U8_F32. + // The extra min is required to pattern match to V_CVT_PKACCUM_U8_F32. + AU1 ABuc0ToU1(AU1 d,AF1 i){return (d&0xffffff00u)|((min(AU1(i),255u) )&(0x000000ffu));} + AU1 ABuc1ToU1(AU1 d,AF1 i){return (d&0xffff00ffu)|((min(AU1(i),255u)<< 8)&(0x0000ff00u));} + AU1 ABuc2ToU1(AU1 d,AF1 i){return (d&0xff00ffffu)|((min(AU1(i),255u)<<16)&(0x00ff0000u));} + AU1 ABuc3ToU1(AU1 d,AF1 i){return (d&0x00ffffffu)|((min(AU1(i),255u)<<24)&(0xff000000u));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed to be one V_CVT_F32_UBYTE*. + AF1 ABuc0FromU1(AU1 i){return AF1((i )&255u);} + AF1 ABuc1FromU1(AU1 i){return AF1((i>> 8)&255u);} + AF1 ABuc2FromU1(AU1 i){return AF1((i>>16)&255u);} + AF1 ABuc3FromU1(AU1 i){return AF1((i>>24)&255u);} + #endif +//============================================================================================================================== + #ifdef A_HALF + // Takes {x0,x1} and {y0,y1} and builds {{x0,y0},{x1,y1}}. + AW2 ABuc01ToW2(AH2 x,AH2 y){x*=AH2_(1.0/32768.0);y*=AH2_(1.0/32768.0); + return AW2_AU1(APermGCEA(AU2(AU1_AW2(AW2_AH2(x)),AU1_AW2(AW2_AH2(y)))));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed for 3 ops to do SOA to AOS and conversion. + AU2 ABuc0ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABuc1ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABuc2ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABuc3ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed for 2 ops to do both AOS to SOA, and conversion. + AH2 ABuc0FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)))*AH2_(32768.0);} + AH2 ABuc1FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)))*AH2_(32768.0);} + AH2 ABuc2FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)))*AH2_(32768.0);} + AH2 ABuc3FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)))*AH2_(32768.0);} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [BSC] BYTE SIGNED CONVERSION +//------------------------------------------------------------------------------------------------------------------------------ +// Similar to [BUC]. +// Works on a range of {-/+ A_BSC_<32,16>}, for <32-bit, and 16-bit> respectively. +//------------------------------------------------------------------------------------------------------------------------------ +// ENCODING (without zero-based encoding) +// ======== +// 0 = unused (can be used to mean something else) +// 1 = lowest value +// 128 = exact zero center (zero based encoding +// 255 = highest value +//------------------------------------------------------------------------------------------------------------------------------ +// Zero-based [Zb] flips the MSB bit of the byte (making 128 "exact zero" actually zero). +// This is useful if there is a desire for cleared values to decode as zero. +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABsc{0,1,2,3}{To,From}U2() - Designed for 16-bit denormal tricks and V_PERM_B32. +// ==== ===== +// 0 : -127/512 (unused) +// 1 : -126/512 +// 2 : -125/512 +// ... +// 128 : 0 +// ... +// 255 : 127/512 +// : 1/4 (just outside the encoding range) +//============================================================================================================================== + // Peak range for 32-bit and 16-bit operations. + #define A_BSC_32 (127.0) + #define A_BSC_16 (127.0/512.0) +//============================================================================================================================== + #if 1 + AU1 ABsc0ToU1(AU1 d,AF1 i){return (d&0xffffff00u)|((min(AU1(i+128.0),255u) )&(0x000000ffu));} + AU1 ABsc1ToU1(AU1 d,AF1 i){return (d&0xffff00ffu)|((min(AU1(i+128.0),255u)<< 8)&(0x0000ff00u));} + AU1 ABsc2ToU1(AU1 d,AF1 i){return (d&0xff00ffffu)|((min(AU1(i+128.0),255u)<<16)&(0x00ff0000u));} + AU1 ABsc3ToU1(AU1 d,AF1 i){return (d&0x00ffffffu)|((min(AU1(i+128.0),255u)<<24)&(0xff000000u));} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 ABsc0ToZbU1(AU1 d,AF1 i){return ((d&0xffffff00u)|((min(AU1(trunc(i)+128.0),255u) )&(0x000000ffu)))^0x00000080u;} + AU1 ABsc1ToZbU1(AU1 d,AF1 i){return ((d&0xffff00ffu)|((min(AU1(trunc(i)+128.0),255u)<< 8)&(0x0000ff00u)))^0x00008000u;} + AU1 ABsc2ToZbU1(AU1 d,AF1 i){return ((d&0xff00ffffu)|((min(AU1(trunc(i)+128.0),255u)<<16)&(0x00ff0000u)))^0x00800000u;} + AU1 ABsc3ToZbU1(AU1 d,AF1 i){return ((d&0x00ffffffu)|((min(AU1(trunc(i)+128.0),255u)<<24)&(0xff000000u)))^0x80000000u;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ABsc0FromU1(AU1 i){return AF1((i )&255u)-128.0;} + AF1 ABsc1FromU1(AU1 i){return AF1((i>> 8)&255u)-128.0;} + AF1 ABsc2FromU1(AU1 i){return AF1((i>>16)&255u)-128.0;} + AF1 ABsc3FromU1(AU1 i){return AF1((i>>24)&255u)-128.0;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ABsc0FromZbU1(AU1 i){return AF1(((i )&255u)^0x80u)-128.0;} + AF1 ABsc1FromZbU1(AU1 i){return AF1(((i>> 8)&255u)^0x80u)-128.0;} + AF1 ABsc2FromZbU1(AU1 i){return AF1(((i>>16)&255u)^0x80u)-128.0;} + AF1 ABsc3FromZbU1(AU1 i){return AF1(((i>>24)&255u)^0x80u)-128.0;} + #endif +//============================================================================================================================== + #ifdef A_HALF + // Takes {x0,x1} and {y0,y1} and builds {{x0,y0},{x1,y1}}. + AW2 ABsc01ToW2(AH2 x,AH2 y){x=x*AH2_(1.0/32768.0)+AH2_(0.25/32768.0);y=y*AH2_(1.0/32768.0)+AH2_(0.25/32768.0); + return AW2_AU1(APermGCEA(AU2(AU1_AW2(AW2_AH2(x)),AU1_AW2(AW2_AH2(y)))));} +//------------------------------------------------------------------------------------------------------------------------------ + AU2 ABsc0ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABsc1ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABsc2ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABsc3ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AU2 ABsc0ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABsc1ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABsc2ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABsc3ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH2 ABsc0FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc1FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc2FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc3FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)))*AH2_(32768.0)-AH2_(0.25);} +//------------------------------------------------------------------------------------------------------------------------------ + AH2 ABsc0FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc1FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc2FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc3FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// HALF APPROXIMATIONS +//------------------------------------------------------------------------------------------------------------------------------ +// These support only positive inputs. +// Did not see value yet in specialization for range. +// Using quick testing, ended up mostly getting the same "best" approximation for various ranges. +// With hardware that can co-execute transcendentals, the value in approximations could be less than expected. +// However from a latency perspective, if execution of a transcendental is 4 clk, with no packed support, -> 8 clk total. +// And co-execution would require a compiler interleaving a lot of independent work for packed usage. +//------------------------------------------------------------------------------------------------------------------------------ +// The one Newton Raphson iteration form of rsq() was skipped (requires 6 ops total). +// Same with sqrt(), as this could be x*rsq() (7 ops). +//============================================================================================================================== + #ifdef A_HALF + // Minimize squared error across full positive range, 2 ops. + // The 0x1de2 based approximation maps {0 to 1} input maps to < 1 output. + AH1 APrxLoSqrtH1(AH1 a){return AH1_AW1((AW1_AH1(a)>>AW1_(1))+AW1_(0x1de2));} + AH2 APrxLoSqrtH2(AH2 a){return AH2_AW2((AW2_AH2(a)>>AW2_(1))+AW2_(0x1de2));} + AH3 APrxLoSqrtH3(AH3 a){return AH3_AW3((AW3_AH3(a)>>AW3_(1))+AW3_(0x1de2));} + AH4 APrxLoSqrtH4(AH4 a){return AH4_AW4((AW4_AH4(a)>>AW4_(1))+AW4_(0x1de2));} +//------------------------------------------------------------------------------------------------------------------------------ + // Lower precision estimation, 1 op. + // Minimize squared error across {smallest normal to 16384.0}. + AH1 APrxLoRcpH1(AH1 a){return AH1_AW1(AW1_(0x7784)-AW1_AH1(a));} + AH2 APrxLoRcpH2(AH2 a){return AH2_AW2(AW2_(0x7784)-AW2_AH2(a));} + AH3 APrxLoRcpH3(AH3 a){return AH3_AW3(AW3_(0x7784)-AW3_AH3(a));} + AH4 APrxLoRcpH4(AH4 a){return AH4_AW4(AW4_(0x7784)-AW4_AH4(a));} +//------------------------------------------------------------------------------------------------------------------------------ + // Medium precision estimation, one Newton Raphson iteration, 3 ops. + AH1 APrxMedRcpH1(AH1 a){AH1 b=AH1_AW1(AW1_(0x778d)-AW1_AH1(a));return b*(-b*a+AH1_(2.0));} + AH2 APrxMedRcpH2(AH2 a){AH2 b=AH2_AW2(AW2_(0x778d)-AW2_AH2(a));return b*(-b*a+AH2_(2.0));} + AH3 APrxMedRcpH3(AH3 a){AH3 b=AH3_AW3(AW3_(0x778d)-AW3_AH3(a));return b*(-b*a+AH3_(2.0));} + AH4 APrxMedRcpH4(AH4 a){AH4 b=AH4_AW4(AW4_(0x778d)-AW4_AH4(a));return b*(-b*a+AH4_(2.0));} +//------------------------------------------------------------------------------------------------------------------------------ + // Minimize squared error across {smallest normal to 16384.0}, 2 ops. + AH1 APrxLoRsqH1(AH1 a){return AH1_AW1(AW1_(0x59a3)-(AW1_AH1(a)>>AW1_(1)));} + AH2 APrxLoRsqH2(AH2 a){return AH2_AW2(AW2_(0x59a3)-(AW2_AH2(a)>>AW2_(1)));} + AH3 APrxLoRsqH3(AH3 a){return AH3_AW3(AW3_(0x59a3)-(AW3_AH3(a)>>AW3_(1)));} + AH4 APrxLoRsqH4(AH4 a){return AH4_AW4(AW4_(0x59a3)-(AW4_AH4(a)>>AW4_(1)));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// FLOAT APPROXIMATIONS +//------------------------------------------------------------------------------------------------------------------------------ +// Michal Drobot has an excellent presentation on these: "Low Level Optimizations For GCN", +// - Idea dates back to SGI, then to Quake 3, etc. +// - https://michaldrobot.files.wordpress.com/2014/05/gcn_alu_opt_digitaldragons2014.pdf +// - sqrt(x)=rsqrt(x)*x +// - rcp(x)=rsqrt(x)*rsqrt(x) for positive x +// - https://github.com/michaldrobot/ShaderFastLibs/blob/master/ShaderFastMathLib.h +//------------------------------------------------------------------------------------------------------------------------------ +// These below are from perhaps less complete searching for optimal. +// Used FP16 normal range for testing with +4096 32-bit step size for sampling error. +// So these match up well with the half approximations. +//============================================================================================================================== + AF1 APrxLoSqrtF1(AF1 a){return AF1_AU1((AU1_AF1(a)>>AU1_(1))+AU1_(0x1fbc4639));} + AF1 APrxLoRcpF1(AF1 a){return AF1_AU1(AU1_(0x7ef07ebb)-AU1_AF1(a));} + AF1 APrxMedRcpF1(AF1 a){AF1 b=AF1_AU1(AU1_(0x7ef19fff)-AU1_AF1(a));return b*(-b*a+AF1_(2.0));} + AF1 APrxLoRsqF1(AF1 a){return AF1_AU1(AU1_(0x5f347d74)-(AU1_AF1(a)>>AU1_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 APrxLoSqrtF2(AF2 a){return AF2_AU2((AU2_AF2(a)>>AU2_(1))+AU2_(0x1fbc4639));} + AF2 APrxLoRcpF2(AF2 a){return AF2_AU2(AU2_(0x7ef07ebb)-AU2_AF2(a));} + AF2 APrxMedRcpF2(AF2 a){AF2 b=AF2_AU2(AU2_(0x7ef19fff)-AU2_AF2(a));return b*(-b*a+AF2_(2.0));} + AF2 APrxLoRsqF2(AF2 a){return AF2_AU2(AU2_(0x5f347d74)-(AU2_AF2(a)>>AU2_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF3 APrxLoSqrtF3(AF3 a){return AF3_AU3((AU3_AF3(a)>>AU3_(1))+AU3_(0x1fbc4639));} + AF3 APrxLoRcpF3(AF3 a){return AF3_AU3(AU3_(0x7ef07ebb)-AU3_AF3(a));} + AF3 APrxMedRcpF3(AF3 a){AF3 b=AF3_AU3(AU3_(0x7ef19fff)-AU3_AF3(a));return b*(-b*a+AF3_(2.0));} + AF3 APrxLoRsqF3(AF3 a){return AF3_AU3(AU3_(0x5f347d74)-(AU3_AF3(a)>>AU3_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF4 APrxLoSqrtF4(AF4 a){return AF4_AU4((AU4_AF4(a)>>AU4_(1))+AU4_(0x1fbc4639));} + AF4 APrxLoRcpF4(AF4 a){return AF4_AU4(AU4_(0x7ef07ebb)-AU4_AF4(a));} + AF4 APrxMedRcpF4(AF4 a){AF4 b=AF4_AU4(AU4_(0x7ef19fff)-AU4_AF4(a));return b*(-b*a+AF4_(2.0));} + AF4 APrxLoRsqF4(AF4 a){return AF4_AU4(AU4_(0x5f347d74)-(AU4_AF4(a)>>AU4_(1)));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PQ APPROXIMATIONS +//------------------------------------------------------------------------------------------------------------------------------ +// PQ is very close to x^(1/8). The functions below Use the fast float approximation method to do +// PQ<~>Gamma2 (4th power and fast 4th root) and PQ<~>Linear (8th power and fast 8th root). Maximum error is ~0.2%. +//============================================================================================================================== +// Helpers + AF1 Quart(AF1 a) { a = a * a; return a * a;} + AF1 Oct(AF1 a) { a = a * a; a = a * a; return a * a; } + AF2 Quart(AF2 a) { a = a * a; return a * a; } + AF2 Oct(AF2 a) { a = a * a; a = a * a; return a * a; } + AF3 Quart(AF3 a) { a = a * a; return a * a; } + AF3 Oct(AF3 a) { a = a * a; a = a * a; return a * a; } + AF4 Quart(AF4 a) { a = a * a; return a * a; } + AF4 Oct(AF4 a) { a = a * a; a = a * a; return a * a; } + //------------------------------------------------------------------------------------------------------------------------------ + AF1 APrxPQToGamma2(AF1 a) { return Quart(a); } + AF1 APrxPQToLinear(AF1 a) { return Oct(a); } + AF1 APrxLoGamma2ToPQ(AF1 a) { return AF1_AU1((AU1_AF1(a) >> AU1_(2)) + AU1_(0x2F9A4E46)); } + AF1 APrxMedGamma2ToPQ(AF1 a) { AF1 b = AF1_AU1((AU1_AF1(a) >> AU1_(2)) + AU1_(0x2F9A4E46)); AF1 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF1 APrxHighGamma2ToPQ(AF1 a) { return sqrt(sqrt(a)); } + AF1 APrxLoLinearToPQ(AF1 a) { return AF1_AU1((AU1_AF1(a) >> AU1_(3)) + AU1_(0x378D8723)); } + AF1 APrxMedLinearToPQ(AF1 a) { AF1 b = AF1_AU1((AU1_AF1(a) >> AU1_(3)) + AU1_(0x378D8723)); AF1 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF1 APrxHighLinearToPQ(AF1 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF2 APrxPQToGamma2(AF2 a) { return Quart(a); } + AF2 APrxPQToLinear(AF2 a) { return Oct(a); } + AF2 APrxLoGamma2ToPQ(AF2 a) { return AF2_AU2((AU2_AF2(a) >> AU2_(2)) + AU2_(0x2F9A4E46)); } + AF2 APrxMedGamma2ToPQ(AF2 a) { AF2 b = AF2_AU2((AU2_AF2(a) >> AU2_(2)) + AU2_(0x2F9A4E46)); AF2 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF2 APrxHighGamma2ToPQ(AF2 a) { return sqrt(sqrt(a)); } + AF2 APrxLoLinearToPQ(AF2 a) { return AF2_AU2((AU2_AF2(a) >> AU2_(3)) + AU2_(0x378D8723)); } + AF2 APrxMedLinearToPQ(AF2 a) { AF2 b = AF2_AU2((AU2_AF2(a) >> AU2_(3)) + AU2_(0x378D8723)); AF2 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF2 APrxHighLinearToPQ(AF2 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF3 APrxPQToGamma2(AF3 a) { return Quart(a); } + AF3 APrxPQToLinear(AF3 a) { return Oct(a); } + AF3 APrxLoGamma2ToPQ(AF3 a) { return AF3_AU3((AU3_AF3(a) >> AU3_(2)) + AU3_(0x2F9A4E46)); } + AF3 APrxMedGamma2ToPQ(AF3 a) { AF3 b = AF3_AU3((AU3_AF3(a) >> AU3_(2)) + AU3_(0x2F9A4E46)); AF3 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF3 APrxHighGamma2ToPQ(AF3 a) { return sqrt(sqrt(a)); } + AF3 APrxLoLinearToPQ(AF3 a) { return AF3_AU3((AU3_AF3(a) >> AU3_(3)) + AU3_(0x378D8723)); } + AF3 APrxMedLinearToPQ(AF3 a) { AF3 b = AF3_AU3((AU3_AF3(a) >> AU3_(3)) + AU3_(0x378D8723)); AF3 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF3 APrxHighLinearToPQ(AF3 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF4 APrxPQToGamma2(AF4 a) { return Quart(a); } + AF4 APrxPQToLinear(AF4 a) { return Oct(a); } + AF4 APrxLoGamma2ToPQ(AF4 a) { return AF4_AU4((AU4_AF4(a) >> AU4_(2)) + AU4_(0x2F9A4E46)); } + AF4 APrxMedGamma2ToPQ(AF4 a) { AF4 b = AF4_AU4((AU4_AF4(a) >> AU4_(2)) + AU4_(0x2F9A4E46)); AF4 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF4 APrxHighGamma2ToPQ(AF4 a) { return sqrt(sqrt(a)); } + AF4 APrxLoLinearToPQ(AF4 a) { return AF4_AU4((AU4_AF4(a) >> AU4_(3)) + AU4_(0x378D8723)); } + AF4 APrxMedLinearToPQ(AF4 a) { AF4 b = AF4_AU4((AU4_AF4(a) >> AU4_(3)) + AU4_(0x378D8723)); AF4 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF4 APrxHighLinearToPQ(AF4 a) { return sqrt(sqrt(sqrt(a))); } +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PARABOLIC SIN & COS +//------------------------------------------------------------------------------------------------------------------------------ +// Approximate answers to transcendental questions. +//------------------------------------------------------------------------------------------------------------------------------ +//============================================================================================================================== + #if 1 + // Valid input range is {-1 to 1} representing {0 to 2 pi}. + // Output range is {-1/4 to 1/4} representing {-1 to 1}. + AF1 APSinF1(AF1 x){return x*abs(x)-x;} // MAD. + AF2 APSinF2(AF2 x){return x*abs(x)-x;} + AF1 APCosF1(AF1 x){x=AFractF1(x*AF1_(0.5)+AF1_(0.75));x=x*AF1_(2.0)-AF1_(1.0);return APSinF1(x);} // 3x MAD, FRACT + AF2 APCosF2(AF2 x){x=AFractF2(x*AF2_(0.5)+AF2_(0.75));x=x*AF2_(2.0)-AF2_(1.0);return APSinF2(x);} + AF2 APSinCosF1(AF1 x){AF1 y=AFractF1(x*AF1_(0.5)+AF1_(0.75));y=y*AF1_(2.0)-AF1_(1.0);return APSinF2(AF2(x,y));} + #endif +//------------------------------------------------------------------------------------------------------------------------------ + #ifdef A_HALF + // For a packed {sin,cos} pair, + // - Native takes 16 clocks and 4 issue slots (no packed transcendentals). + // - Parabolic takes 8 clocks and 8 issue slots (only fract is non-packed). + AH1 APSinH1(AH1 x){return x*abs(x)-x;} + AH2 APSinH2(AH2 x){return x*abs(x)-x;} // AND,FMA + AH1 APCosH1(AH1 x){x=AFractH1(x*AH1_(0.5)+AH1_(0.75));x=x*AH1_(2.0)-AH1_(1.0);return APSinH1(x);} + AH2 APCosH2(AH2 x){x=AFractH2(x*AH2_(0.5)+AH2_(0.75));x=x*AH2_(2.0)-AH2_(1.0);return APSinH2(x);} // 3x FMA, 2xFRACT, AND + AH2 APSinCosH1(AH1 x){AH1 y=AFractH1(x*AH1_(0.5)+AH1_(0.75));y=y*AH1_(2.0)-AH1_(1.0);return APSinH2(AH2(x,y));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [ZOL] ZERO ONE LOGIC +//------------------------------------------------------------------------------------------------------------------------------ +// Conditional free logic designed for easy 16-bit packing, and backwards porting to 32-bit. +//------------------------------------------------------------------------------------------------------------------------------ +// 0 := false +// 1 := true +//------------------------------------------------------------------------------------------------------------------------------ +// AndNot(x,y) -> !(x&y) .... One op. +// AndOr(x,y,z) -> (x&y)|z ... One op. +// GtZero(x) -> x>0.0 ..... One op. +// Sel(x,y,z) -> x?y:z ..... Two ops, has no precision loss. +// Signed(x) -> x<0.0 ..... One op. +// ZeroPass(x,y) -> x?0:y ..... Two ops, 'y' is a pass through safe for aliasing as integer. +//------------------------------------------------------------------------------------------------------------------------------ +// OPTIMIZATION NOTES +// ================== +// - On Vega to use 2 constants in a packed op, pass in as one AW2 or one AH2 'k.xy' and use as 'k.xx' and 'k.yy'. +// For example 'a.xy*k.xx+k.yy'. +//============================================================================================================================== + #if 1 + AU1 AZolAndU1(AU1 x,AU1 y){return min(x,y);} + AU2 AZolAndU2(AU2 x,AU2 y){return min(x,y);} + AU3 AZolAndU3(AU3 x,AU3 y){return min(x,y);} + AU4 AZolAndU4(AU4 x,AU4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AZolNotU1(AU1 x){return x^AU1_(1);} + AU2 AZolNotU2(AU2 x){return x^AU2_(1);} + AU3 AZolNotU3(AU3 x){return x^AU3_(1);} + AU4 AZolNotU4(AU4 x){return x^AU4_(1);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AZolOrU1(AU1 x,AU1 y){return max(x,y);} + AU2 AZolOrU2(AU2 x,AU2 y){return max(x,y);} + AU3 AZolOrU3(AU3 x,AU3 y){return max(x,y);} + AU4 AZolOrU4(AU4 x,AU4 y){return max(x,y);} +//============================================================================================================================== + AU1 AZolF1ToU1(AF1 x){return AU1(x);} + AU2 AZolF2ToU2(AF2 x){return AU2(x);} + AU3 AZolF3ToU3(AF3 x){return AU3(x);} + AU4 AZolF4ToU4(AF4 x){return AU4(x);} +//------------------------------------------------------------------------------------------------------------------------------ + // 2 ops, denormals don't work in 32-bit on PC (and if they are enabled, OMOD is disabled). + AU1 AZolNotF1ToU1(AF1 x){return AU1(AF1_(1.0)-x);} + AU2 AZolNotF2ToU2(AF2 x){return AU2(AF2_(1.0)-x);} + AU3 AZolNotF3ToU3(AF3 x){return AU3(AF3_(1.0)-x);} + AU4 AZolNotF4ToU4(AF4 x){return AU4(AF4_(1.0)-x);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolU1ToF1(AU1 x){return AF1(x);} + AF2 AZolU2ToF2(AU2 x){return AF2(x);} + AF3 AZolU3ToF3(AU3 x){return AF3(x);} + AF4 AZolU4ToF4(AU4 x){return AF4(x);} +//============================================================================================================================== + AF1 AZolAndF1(AF1 x,AF1 y){return min(x,y);} + AF2 AZolAndF2(AF2 x,AF2 y){return min(x,y);} + AF3 AZolAndF3(AF3 x,AF3 y){return min(x,y);} + AF4 AZolAndF4(AF4 x,AF4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ASolAndNotF1(AF1 x,AF1 y){return (-x)*y+AF1_(1.0);} + AF2 ASolAndNotF2(AF2 x,AF2 y){return (-x)*y+AF2_(1.0);} + AF3 ASolAndNotF3(AF3 x,AF3 y){return (-x)*y+AF3_(1.0);} + AF4 ASolAndNotF4(AF4 x,AF4 y){return (-x)*y+AF4_(1.0);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolAndOrF1(AF1 x,AF1 y,AF1 z){return ASatF1(x*y+z);} + AF2 AZolAndOrF2(AF2 x,AF2 y,AF2 z){return ASatF2(x*y+z);} + AF3 AZolAndOrF3(AF3 x,AF3 y,AF3 z){return ASatF3(x*y+z);} + AF4 AZolAndOrF4(AF4 x,AF4 y,AF4 z){return ASatF4(x*y+z);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolGtZeroF1(AF1 x){return ASatF1(x*AF1_(A_INFP_F));} + AF2 AZolGtZeroF2(AF2 x){return ASatF2(x*AF2_(A_INFP_F));} + AF3 AZolGtZeroF3(AF3 x){return ASatF3(x*AF3_(A_INFP_F));} + AF4 AZolGtZeroF4(AF4 x){return ASatF4(x*AF4_(A_INFP_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolNotF1(AF1 x){return AF1_(1.0)-x;} + AF2 AZolNotF2(AF2 x){return AF2_(1.0)-x;} + AF3 AZolNotF3(AF3 x){return AF3_(1.0)-x;} + AF4 AZolNotF4(AF4 x){return AF4_(1.0)-x;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolOrF1(AF1 x,AF1 y){return max(x,y);} + AF2 AZolOrF2(AF2 x,AF2 y){return max(x,y);} + AF3 AZolOrF3(AF3 x,AF3 y){return max(x,y);} + AF4 AZolOrF4(AF4 x,AF4 y){return max(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolSelF1(AF1 x,AF1 y,AF1 z){AF1 r=(-x)*z+z;return x*y+r;} + AF2 AZolSelF2(AF2 x,AF2 y,AF2 z){AF2 r=(-x)*z+z;return x*y+r;} + AF3 AZolSelF3(AF3 x,AF3 y,AF3 z){AF3 r=(-x)*z+z;return x*y+r;} + AF4 AZolSelF4(AF4 x,AF4 y,AF4 z){AF4 r=(-x)*z+z;return x*y+r;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolSignedF1(AF1 x){return ASatF1(x*AF1_(A_INFN_F));} + AF2 AZolSignedF2(AF2 x){return ASatF2(x*AF2_(A_INFN_F));} + AF3 AZolSignedF3(AF3 x){return ASatF3(x*AF3_(A_INFN_F));} + AF4 AZolSignedF4(AF4 x){return ASatF4(x*AF4_(A_INFN_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolZeroPassF1(AF1 x,AF1 y){return AF1_AU1((AU1_AF1(x)!=AU1_(0))?AU1_(0):AU1_AF1(y));} + AF2 AZolZeroPassF2(AF2 x,AF2 y){return AF2_AU2((AU2_AF2(x)!=AU2_(0))?AU2_(0):AU2_AF2(y));} + AF3 AZolZeroPassF3(AF3 x,AF3 y){return AF3_AU3((AU3_AF3(x)!=AU3_(0))?AU3_(0):AU3_AF3(y));} + AF4 AZolZeroPassF4(AF4 x,AF4 y){return AF4_AU4((AU4_AF4(x)!=AU4_(0))?AU4_(0):AU4_AF4(y));} + #endif +//============================================================================================================================== + #ifdef A_HALF + AW1 AZolAndW1(AW1 x,AW1 y){return min(x,y);} + AW2 AZolAndW2(AW2 x,AW2 y){return min(x,y);} + AW3 AZolAndW3(AW3 x,AW3 y){return min(x,y);} + AW4 AZolAndW4(AW4 x,AW4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AZolNotW1(AW1 x){return x^AW1_(1);} + AW2 AZolNotW2(AW2 x){return x^AW2_(1);} + AW3 AZolNotW3(AW3 x){return x^AW3_(1);} + AW4 AZolNotW4(AW4 x){return x^AW4_(1);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AZolOrW1(AW1 x,AW1 y){return max(x,y);} + AW2 AZolOrW2(AW2 x,AW2 y){return max(x,y);} + AW3 AZolOrW3(AW3 x,AW3 y){return max(x,y);} + AW4 AZolOrW4(AW4 x,AW4 y){return max(x,y);} +//============================================================================================================================== + // Uses denormal trick. + AW1 AZolH1ToW1(AH1 x){return AW1_AH1(x*AH1_AW1(AW1_(1)));} + AW2 AZolH2ToW2(AH2 x){return AW2_AH2(x*AH2_AW2(AW2_(1)));} + AW3 AZolH3ToW3(AH3 x){return AW3_AH3(x*AH3_AW3(AW3_(1)));} + AW4 AZolH4ToW4(AH4 x){return AW4_AH4(x*AH4_AW4(AW4_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + // AMD arch lacks a packed conversion opcode. + AH1 AZolW1ToH1(AW1 x){return AH1_AW1(x*AW1_AH1(AH1_(1.0)));} + AH2 AZolW2ToH2(AW2 x){return AH2_AW2(x*AW2_AH2(AH2_(1.0)));} + AH3 AZolW1ToH3(AW3 x){return AH3_AW3(x*AW3_AH3(AH3_(1.0)));} + AH4 AZolW2ToH4(AW4 x){return AH4_AW4(x*AW4_AH4(AH4_(1.0)));} +//============================================================================================================================== + AH1 AZolAndH1(AH1 x,AH1 y){return min(x,y);} + AH2 AZolAndH2(AH2 x,AH2 y){return min(x,y);} + AH3 AZolAndH3(AH3 x,AH3 y){return min(x,y);} + AH4 AZolAndH4(AH4 x,AH4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASolAndNotH1(AH1 x,AH1 y){return (-x)*y+AH1_(1.0);} + AH2 ASolAndNotH2(AH2 x,AH2 y){return (-x)*y+AH2_(1.0);} + AH3 ASolAndNotH3(AH3 x,AH3 y){return (-x)*y+AH3_(1.0);} + AH4 ASolAndNotH4(AH4 x,AH4 y){return (-x)*y+AH4_(1.0);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolAndOrH1(AH1 x,AH1 y,AH1 z){return ASatH1(x*y+z);} + AH2 AZolAndOrH2(AH2 x,AH2 y,AH2 z){return ASatH2(x*y+z);} + AH3 AZolAndOrH3(AH3 x,AH3 y,AH3 z){return ASatH3(x*y+z);} + AH4 AZolAndOrH4(AH4 x,AH4 y,AH4 z){return ASatH4(x*y+z);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolGtZeroH1(AH1 x){return ASatH1(x*AH1_(A_INFP_H));} + AH2 AZolGtZeroH2(AH2 x){return ASatH2(x*AH2_(A_INFP_H));} + AH3 AZolGtZeroH3(AH3 x){return ASatH3(x*AH3_(A_INFP_H));} + AH4 AZolGtZeroH4(AH4 x){return ASatH4(x*AH4_(A_INFP_H));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolNotH1(AH1 x){return AH1_(1.0)-x;} + AH2 AZolNotH2(AH2 x){return AH2_(1.0)-x;} + AH3 AZolNotH3(AH3 x){return AH3_(1.0)-x;} + AH4 AZolNotH4(AH4 x){return AH4_(1.0)-x;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolOrH1(AH1 x,AH1 y){return max(x,y);} + AH2 AZolOrH2(AH2 x,AH2 y){return max(x,y);} + AH3 AZolOrH3(AH3 x,AH3 y){return max(x,y);} + AH4 AZolOrH4(AH4 x,AH4 y){return max(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolSelH1(AH1 x,AH1 y,AH1 z){AH1 r=(-x)*z+z;return x*y+r;} + AH2 AZolSelH2(AH2 x,AH2 y,AH2 z){AH2 r=(-x)*z+z;return x*y+r;} + AH3 AZolSelH3(AH3 x,AH3 y,AH3 z){AH3 r=(-x)*z+z;return x*y+r;} + AH4 AZolSelH4(AH4 x,AH4 y,AH4 z){AH4 r=(-x)*z+z;return x*y+r;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolSignedH1(AH1 x){return ASatH1(x*AH1_(A_INFN_H));} + AH2 AZolSignedH2(AH2 x){return ASatH2(x*AH2_(A_INFN_H));} + AH3 AZolSignedH3(AH3 x){return ASatH3(x*AH3_(A_INFN_H));} + AH4 AZolSignedH4(AH4 x){return ASatH4(x*AH4_(A_INFN_H));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// COLOR CONVERSIONS +//------------------------------------------------------------------------------------------------------------------------------ +// These are all linear to/from some other space (where 'linear' has been shortened out of the function name). +// So 'ToGamma' is 'LinearToGamma', and 'FromGamma' is 'LinearFromGamma'. +// These are branch free implementations. +// The AToSrgbF1() function is useful for stores for compute shaders for GPUs without hardware linear->sRGB store conversion. +//------------------------------------------------------------------------------------------------------------------------------ +// TRANSFER FUNCTIONS +// ================== +// 709 ..... Rec709 used for some HDTVs +// Gamma ... Typically 2.2 for some PC displays, or 2.4-2.5 for CRTs, or 2.2 FreeSync2 native +// Pq ...... PQ native for HDR10 +// Srgb .... The sRGB output, typical of PC displays, useful for 10-bit output, or storing to 8-bit UNORM without SRGB type +// Two ..... Gamma 2.0, fastest conversion (useful for intermediate pass approximations) +// Three ... Gamma 3.0, less fast, but good for HDR. +//------------------------------------------------------------------------------------------------------------------------------ +// KEEPING TO SPEC +// =============== +// Both Rec.709 and sRGB have a linear segment which as spec'ed would intersect the curved segment 2 times. +// (a.) For 8-bit sRGB, steps {0 to 10.3} are in the linear region (4% of the encoding range). +// (b.) For 8-bit 709, steps {0 to 20.7} are in the linear region (8% of the encoding range). +// Also there is a slight step in the transition regions. +// Precision of the coefficients in the spec being the likely cause. +// Main usage case of the sRGB code is to do the linear->sRGB converstion in a compute shader before store. +// This is to work around lack of hardware (typically only ROP does the conversion for free). +// To "correct" the linear segment, would be to introduce error, because hardware decode of sRGB->linear is fixed (and free). +// So this header keeps with the spec. +// For linear->sRGB transforms, the linear segment in some respects reduces error, because rounding in that region is linear. +// Rounding in the curved region in hardware (and fast software code) introduces error due to rounding in non-linear. +//------------------------------------------------------------------------------------------------------------------------------ +// FOR PQ +// ====== +// Both input and output is {0.0-1.0}, and where output 1.0 represents 10000.0 cd/m^2. +// All constants are only specified to FP32 precision. +// External PQ source reference, +// - https://github.com/ampas/aces-dev/blob/master/transforms/ctl/utilities/ACESlib.Utilities_Color.a1.0.1.ctl +//------------------------------------------------------------------------------------------------------------------------------ +// PACKED VERSIONS +// =============== +// These are the A*H2() functions. +// There is no PQ functions as FP16 seemed to not have enough precision for the conversion. +// The remaining functions are "good enough" for 8-bit, and maybe 10-bit if not concerned about a few 1-bit errors. +// Precision is lowest in the 709 conversion, higher in sRGB, higher still in Two and Gamma (when using 2.2 at least). +//------------------------------------------------------------------------------------------------------------------------------ +// NOTES +// ===== +// Could be faster for PQ conversions to be in ALU or a texture lookup depending on usage case. +//============================================================================================================================== + #if 1 + AF1 ATo709F1(AF1 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AF2 ATo709F2(AF2 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AF3 ATo709F3(AF3 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + // Note 'rcpX' is '1/x', where the 'x' is what would be used in AFromGamma(). + AF1 AToGammaF1(AF1 c,AF1 rcpX){return pow(c,AF1_(rcpX));} + AF2 AToGammaF2(AF2 c,AF1 rcpX){return pow(c,AF2_(rcpX));} + AF3 AToGammaF3(AF3 c,AF1 rcpX){return pow(c,AF3_(rcpX));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToPqF1(AF1 x){AF1 p=pow(x,AF1_(0.159302)); + return pow((AF1_(0.835938)+AF1_(18.8516)*p)/(AF1_(1.0)+AF1_(18.6875)*p),AF1_(78.8438));} + AF2 AToPqF1(AF2 x){AF2 p=pow(x,AF2_(0.159302)); + return pow((AF2_(0.835938)+AF2_(18.8516)*p)/(AF2_(1.0)+AF2_(18.6875)*p),AF2_(78.8438));} + AF3 AToPqF1(AF3 x){AF3 p=pow(x,AF3_(0.159302)); + return pow((AF3_(0.835938)+AF3_(18.8516)*p)/(AF3_(1.0)+AF3_(18.6875)*p),AF3_(78.8438));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToSrgbF1(AF1 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AF2 AToSrgbF2(AF2 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AF3 AToSrgbF3(AF3 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToTwoF1(AF1 c){return sqrt(c);} + AF2 AToTwoF2(AF2 c){return sqrt(c);} + AF3 AToTwoF3(AF3 c){return sqrt(c);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToThreeF1(AF1 c){return pow(c,AF1_(1.0/3.0));} + AF2 AToThreeF2(AF2 c){return pow(c,AF2_(1.0/3.0));} + AF3 AToThreeF3(AF3 c){return pow(c,AF3_(1.0/3.0));} + #endif +//============================================================================================================================== + #if 1 + // Unfortunately median won't work here. + AF1 AFrom709F1(AF1 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF1(AZolSignedF1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AF2 AFrom709F2(AF2 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF2(AZolSignedF2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AF3 AFrom709F3(AF3 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF3(AZolSignedF3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromGammaF1(AF1 c,AF1 x){return pow(c,AF1_(x));} + AF2 AFromGammaF2(AF2 c,AF1 x){return pow(c,AF2_(x));} + AF3 AFromGammaF3(AF3 c,AF1 x){return pow(c,AF3_(x));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromPqF1(AF1 x){AF1 p=pow(x,AF1_(0.0126833)); + return pow(ASatF1(p-AF1_(0.835938))/(AF1_(18.8516)-AF1_(18.6875)*p),AF1_(6.27739));} + AF2 AFromPqF1(AF2 x){AF2 p=pow(x,AF2_(0.0126833)); + return pow(ASatF2(p-AF2_(0.835938))/(AF2_(18.8516)-AF2_(18.6875)*p),AF2_(6.27739));} + AF3 AFromPqF1(AF3 x){AF3 p=pow(x,AF3_(0.0126833)); + return pow(ASatF3(p-AF3_(0.835938))/(AF3_(18.8516)-AF3_(18.6875)*p),AF3_(6.27739));} +//------------------------------------------------------------------------------------------------------------------------------ + // Unfortunately median won't work here. + AF1 AFromSrgbF1(AF1 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF1(AZolSignedF1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AF2 AFromSrgbF2(AF2 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF2(AZolSignedF2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AF3 AFromSrgbF3(AF3 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF3(AZolSignedF3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromTwoF1(AF1 c){return c*c;} + AF2 AFromTwoF2(AF2 c){return c*c;} + AF3 AFromTwoF3(AF3 c){return c*c;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AFromThreeF1(AF1 c){return c*c*c;} + AF2 AFromThreeF2(AF2 c){return c*c*c;} + AF3 AFromThreeF3(AF3 c){return c*c*c;} + #endif +//============================================================================================================================== + #ifdef A_HALF + AH1 ATo709H1(AH1 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AH2 ATo709H2(AH2 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AH3 ATo709H3(AH3 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToGammaH1(AH1 c,AH1 rcpX){return pow(c,AH1_(rcpX));} + AH2 AToGammaH2(AH2 c,AH1 rcpX){return pow(c,AH2_(rcpX));} + AH3 AToGammaH3(AH3 c,AH1 rcpX){return pow(c,AH3_(rcpX));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToSrgbH1(AH1 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AH2 AToSrgbH2(AH2 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AH3 AToSrgbH3(AH3 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToTwoH1(AH1 c){return sqrt(c);} + AH2 AToTwoH2(AH2 c){return sqrt(c);} + AH3 AToTwoH3(AH3 c){return sqrt(c);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToThreeF1(AH1 c){return pow(c,AH1_(1.0/3.0));} + AH2 AToThreeF2(AH2 c){return pow(c,AH2_(1.0/3.0));} + AH3 AToThreeF3(AH3 c){return pow(c,AH3_(1.0/3.0));} + #endif +//============================================================================================================================== + #ifdef A_HALF + AH1 AFrom709H1(AH1 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH1(AZolSignedH1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AH2 AFrom709H2(AH2 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH2(AZolSignedH2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AH3 AFrom709H3(AH3 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH3(AZolSignedH3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromGammaH1(AH1 c,AH1 x){return pow(c,AH1_(x));} + AH2 AFromGammaH2(AH2 c,AH1 x){return pow(c,AH2_(x));} + AH3 AFromGammaH3(AH3 c,AH1 x){return pow(c,AH3_(x));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AHromSrgbF1(AH1 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH1(AZolSignedH1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AH2 AHromSrgbF2(AH2 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH2(AZolSignedH2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AH3 AHromSrgbF3(AH3 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH3(AZolSignedH3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromTwoH1(AH1 c){return c*c;} + AH2 AFromTwoH2(AH2 c){return c*c;} + AH3 AFromTwoH3(AH3 c){return c*c;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromThreeH1(AH1 c){return c*c*c;} + AH2 AFromThreeH2(AH2 c){return c*c*c;} + AH3 AFromThreeH3(AH3 c){return c*c*c;} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CS REMAP +//============================================================================================================================== + // Simple remap 64x1 to 8x8 with rotated 2x2 pixel quads in quad linear. + // 543210 + // ====== + // ..xxx. + // yy...y + AU2 ARmp8x8(AU1 a){return AU2(ABfe(a,1u,3u),ABfiM(ABfe(a,3u,3u),a,1u));} +//============================================================================================================================== + // More complex remap 64x1 to 8x8 which is necessary for 2D wave reductions. + // 543210 + // ====== + // .xx..x + // y..yy. + // Details, + // LANE TO 8x8 MAPPING + // =================== + // 00 01 08 09 10 11 18 19 + // 02 03 0a 0b 12 13 1a 1b + // 04 05 0c 0d 14 15 1c 1d + // 06 07 0e 0f 16 17 1e 1f + // 20 21 28 29 30 31 38 39 + // 22 23 2a 2b 32 33 3a 3b + // 24 25 2c 2d 34 35 3c 3d + // 26 27 2e 2f 36 37 3e 3f + AU2 ARmpRed8x8(AU1 a){return AU2(ABfiM(ABfe(a,2u,3u),a,1u),ABfiM(ABfe(a,3u,3u),ABfe(a,1u,2u),2u));} +//============================================================================================================================== + #ifdef A_HALF + AW2 ARmp8x8H(AU1 a){return AW2(ABfe(a,1u,3u),ABfiM(ABfe(a,3u,3u),a,1u));} + AW2 ARmpRed8x8H(AU1 a){return AW2(ABfiM(ABfe(a,2u,3u),a,1u),ABfiM(ABfe(a,3u,3u),ABfe(a,1u,2u),2u));} + #endif +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// REFERENCE +// +//------------------------------------------------------------------------------------------------------------------------------ +// IEEE FLOAT RULES +// ================ +// - saturate(NaN)=0, saturate(-INF)=0, saturate(+INF)=1 +// - {+/-}0 * {+/-}INF = NaN +// - -INF + (+INF) = NaN +// - {+/-}0 / {+/-}0 = NaN +// - {+/-}INF / {+/-}INF = NaN +// - a<(-0) := sqrt(a) = NaN (a=-0.0 won't NaN) +// - 0 == -0 +// - 4/0 = +INF +// - 4/-0 = -INF +// - 4+INF = +INF +// - 4-INF = -INF +// - 4*(+INF) = +INF +// - 4*(-INF) = -INF +// - -4*(+INF) = -INF +// - sqrt(+INF) = +INF +//------------------------------------------------------------------------------------------------------------------------------ +// FP16 ENCODING +// ============= +// fedcba9876543210 +// ---------------- +// ......mmmmmmmmmm 10-bit mantissa (encodes 11-bit 0.5 to 1.0 except for denormals) +// .eeeee.......... 5-bit exponent +// .00000.......... denormals +// .00001.......... -14 exponent +// .11110.......... 15 exponent +// .111110000000000 infinity +// .11111nnnnnnnnnn NaN with n!=0 +// s............... sign +//------------------------------------------------------------------------------------------------------------------------------ +// FP16/INT16 ALIASING DENORMAL +// ============================ +// 11-bit unsigned integers alias with half float denormal/normal values, +// 1 = 2^(-24) = 1/16777216 ....................... first denormal value +// 2 = 2^(-23) +// ... +// 1023 = 2^(-14)*(1-2^(-10)) = 2^(-14)*(1-1/1024) ... last denormal value +// 1024 = 2^(-14) = 1/16384 .......................... first normal value that still maps to integers +// 2047 .............................................. last normal value that still maps to integers +// Scaling limits, +// 2^15 = 32768 ...................................... largest power of 2 scaling +// Largest pow2 conversion mapping is at *32768, +// 1 : 2^(-9) = 1/512 +// 2 : 1/256 +// 4 : 1/128 +// 8 : 1/64 +// 16 : 1/32 +// 32 : 1/16 +// 64 : 1/8 +// 128 : 1/4 +// 256 : 1/2 +// 512 : 1 +// 1024 : 2 +// 2047 : a little less than 4 +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// GPU/CPU PORTABILITY +// +// +//------------------------------------------------------------------------------------------------------------------------------ +// This is the GPU implementation. +// See the CPU implementation for docs. +//============================================================================================================================== +#ifdef A_GPU + #define A_TRUE true + #define A_FALSE false + #define A_STATIC +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR ARGUMENT/RETURN/INITIALIZATION PORTABILITY +//============================================================================================================================== + #define retAD2 AD2 + #define retAD3 AD3 + #define retAD4 AD4 + #define retAF2 AF2 + #define retAF3 AF3 + #define retAF4 AF4 + #define retAL2 AL2 + #define retAL3 AL3 + #define retAL4 AL4 + #define retAU2 AU2 + #define retAU3 AU3 + #define retAU4 AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define inAD2 in AD2 + #define inAD3 in AD3 + #define inAD4 in AD4 + #define inAF2 in AF2 + #define inAF3 in AF3 + #define inAF4 in AF4 + #define inAL2 in AL2 + #define inAL3 in AL3 + #define inAL4 in AL4 + #define inAU2 in AU2 + #define inAU3 in AU3 + #define inAU4 in AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define inoutAD2 inout AD2 + #define inoutAD3 inout AD3 + #define inoutAD4 inout AD4 + #define inoutAF2 inout AF2 + #define inoutAF3 inout AF3 + #define inoutAF4 inout AF4 + #define inoutAL2 inout AL2 + #define inoutAL3 inout AL3 + #define inoutAL4 inout AL4 + #define inoutAU2 inout AU2 + #define inoutAU3 inout AU3 + #define inoutAU4 inout AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define outAD2 out AD2 + #define outAD3 out AD3 + #define outAD4 out AD4 + #define outAF2 out AF2 + #define outAF3 out AF3 + #define outAF4 out AF4 + #define outAL2 out AL2 + #define outAL3 out AL3 + #define outAL4 out AL4 + #define outAU2 out AU2 + #define outAU3 out AU3 + #define outAU4 out AU4 +//------------------------------------------------------------------------------------------------------------------------------ + #define varAD2(x) AD2 x + #define varAD3(x) AD3 x + #define varAD4(x) AD4 x + #define varAF2(x) AF2 x + #define varAF3(x) AF3 x + #define varAF4(x) AF4 x + #define varAL2(x) AL2 x + #define varAL3(x) AL3 x + #define varAL4(x) AL4 x + #define varAU2(x) AU2 x + #define varAU3(x) AU3 x + #define varAU4(x) AU4 x +//------------------------------------------------------------------------------------------------------------------------------ + #define initAD2(x,y) AD2(x,y) + #define initAD3(x,y,z) AD3(x,y,z) + #define initAD4(x,y,z,w) AD4(x,y,z,w) + #define initAF2(x,y) AF2(x,y) + #define initAF3(x,y,z) AF3(x,y,z) + #define initAF4(x,y,z,w) AF4(x,y,z,w) + #define initAL2(x,y) AL2(x,y) + #define initAL3(x,y,z) AL3(x,y,z) + #define initAL4(x,y,z,w) AL4(x,y,z,w) + #define initAU2(x,y) AU2(x,y) + #define initAU3(x,y,z) AU3(x,y,z) + #define initAU4(x,y,z,w) AU4(x,y,z,w) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS +//============================================================================================================================== + #define AAbsD1(a) abs(AD1(a)) + #define AAbsF1(a) abs(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ACosD1(a) cos(AD1(a)) + #define ACosF1(a) cos(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ADotD2(a,b) dot(AD2(a),AD2(b)) + #define ADotD3(a,b) dot(AD3(a),AD3(b)) + #define ADotD4(a,b) dot(AD4(a),AD4(b)) + #define ADotF2(a,b) dot(AF2(a),AF2(b)) + #define ADotF3(a,b) dot(AF3(a),AF3(b)) + #define ADotF4(a,b) dot(AF4(a),AF4(b)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AExp2D1(a) exp2(AD1(a)) + #define AExp2F1(a) exp2(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AFloorD1(a) floor(AD1(a)) + #define AFloorF1(a) floor(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ALog2D1(a) log2(AD1(a)) + #define ALog2F1(a) log2(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define AMaxD1(a,b) max(a,b) + #define AMaxF1(a,b) max(a,b) + #define AMaxL1(a,b) max(a,b) + #define AMaxU1(a,b) max(a,b) +//------------------------------------------------------------------------------------------------------------------------------ + #define AMinD1(a,b) min(a,b) + #define AMinF1(a,b) min(a,b) + #define AMinL1(a,b) min(a,b) + #define AMinU1(a,b) min(a,b) +//------------------------------------------------------------------------------------------------------------------------------ + #define ASinD1(a) sin(AD1(a)) + #define ASinF1(a) sin(AF1(a)) +//------------------------------------------------------------------------------------------------------------------------------ + #define ASqrtD1(a) sqrt(AD1(a)) + #define ASqrtF1(a) sqrt(AF1(a)) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// SCALAR RETURN OPS - DEPENDENT +//============================================================================================================================== + #define APowD1(a,b) pow(AD1(a),AF1(b)) + #define APowF1(a,b) pow(AF1(a),AF1(b)) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// VECTOR OPS +//------------------------------------------------------------------------------------------------------------------------------ +// These are added as needed for production or prototyping, so not necessarily a complete set. +// They follow a convention of taking in a destination and also returning the destination value to increase utility. +//============================================================================================================================== + #ifdef A_DUBL + AD2 opAAbsD2(outAD2 d,inAD2 a){d=abs(a);return d;} + AD3 opAAbsD3(outAD3 d,inAD3 a){d=abs(a);return d;} + AD4 opAAbsD4(outAD4 d,inAD4 a){d=abs(a);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAAddD2(outAD2 d,inAD2 a,inAD2 b){d=a+b;return d;} + AD3 opAAddD3(outAD3 d,inAD3 a,inAD3 b){d=a+b;return d;} + AD4 opAAddD4(outAD4 d,inAD4 a,inAD4 b){d=a+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAAddOneD2(outAD2 d,inAD2 a,AD1 b){d=a+AD2_(b);return d;} + AD3 opAAddOneD3(outAD3 d,inAD3 a,AD1 b){d=a+AD3_(b);return d;} + AD4 opAAddOneD4(outAD4 d,inAD4 a,AD1 b){d=a+AD4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opACpyD2(outAD2 d,inAD2 a){d=a;return d;} + AD3 opACpyD3(outAD3 d,inAD3 a){d=a;return d;} + AD4 opACpyD4(outAD4 d,inAD4 a){d=a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opALerpD2(outAD2 d,inAD2 a,inAD2 b,inAD2 c){d=ALerpD2(a,b,c);return d;} + AD3 opALerpD3(outAD3 d,inAD3 a,inAD3 b,inAD3 c){d=ALerpD3(a,b,c);return d;} + AD4 opALerpD4(outAD4 d,inAD4 a,inAD4 b,inAD4 c){d=ALerpD4(a,b,c);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opALerpOneD2(outAD2 d,inAD2 a,inAD2 b,AD1 c){d=ALerpD2(a,b,AD2_(c));return d;} + AD3 opALerpOneD3(outAD3 d,inAD3 a,inAD3 b,AD1 c){d=ALerpD3(a,b,AD3_(c));return d;} + AD4 opALerpOneD4(outAD4 d,inAD4 a,inAD4 b,AD1 c){d=ALerpD4(a,b,AD4_(c));return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMaxD2(outAD2 d,inAD2 a,inAD2 b){d=max(a,b);return d;} + AD3 opAMaxD3(outAD3 d,inAD3 a,inAD3 b){d=max(a,b);return d;} + AD4 opAMaxD4(outAD4 d,inAD4 a,inAD4 b){d=max(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMinD2(outAD2 d,inAD2 a,inAD2 b){d=min(a,b);return d;} + AD3 opAMinD3(outAD3 d,inAD3 a,inAD3 b){d=min(a,b);return d;} + AD4 opAMinD4(outAD4 d,inAD4 a,inAD4 b){d=min(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMulD2(outAD2 d,inAD2 a,inAD2 b){d=a*b;return d;} + AD3 opAMulD3(outAD3 d,inAD3 a,inAD3 b){d=a*b;return d;} + AD4 opAMulD4(outAD4 d,inAD4 a,inAD4 b){d=a*b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAMulOneD2(outAD2 d,inAD2 a,AD1 b){d=a*AD2_(b);return d;} + AD3 opAMulOneD3(outAD3 d,inAD3 a,AD1 b){d=a*AD3_(b);return d;} + AD4 opAMulOneD4(outAD4 d,inAD4 a,AD1 b){d=a*AD4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opANegD2(outAD2 d,inAD2 a){d=-a;return d;} + AD3 opANegD3(outAD3 d,inAD3 a){d=-a;return d;} + AD4 opANegD4(outAD4 d,inAD4 a){d=-a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opARcpD2(outAD2 d,inAD2 a){d=ARcpD2(a);return d;} + AD3 opARcpD3(outAD3 d,inAD3 a){d=ARcpD3(a);return d;} + AD4 opARcpD4(outAD4 d,inAD4 a){d=ARcpD4(a);return d;} + #endif +//============================================================================================================================== + AF2 opAAbsF2(outAF2 d,inAF2 a){d=abs(a);return d;} + AF3 opAAbsF3(outAF3 d,inAF3 a){d=abs(a);return d;} + AF4 opAAbsF4(outAF4 d,inAF4 a){d=abs(a);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAAddF2(outAF2 d,inAF2 a,inAF2 b){d=a+b;return d;} + AF3 opAAddF3(outAF3 d,inAF3 a,inAF3 b){d=a+b;return d;} + AF4 opAAddF4(outAF4 d,inAF4 a,inAF4 b){d=a+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAAddOneF2(outAF2 d,inAF2 a,AF1 b){d=a+AF2_(b);return d;} + AF3 opAAddOneF3(outAF3 d,inAF3 a,AF1 b){d=a+AF3_(b);return d;} + AF4 opAAddOneF4(outAF4 d,inAF4 a,AF1 b){d=a+AF4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opACpyF2(outAF2 d,inAF2 a){d=a;return d;} + AF3 opACpyF3(outAF3 d,inAF3 a){d=a;return d;} + AF4 opACpyF4(outAF4 d,inAF4 a){d=a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opALerpF2(outAF2 d,inAF2 a,inAF2 b,inAF2 c){d=ALerpF2(a,b,c);return d;} + AF3 opALerpF3(outAF3 d,inAF3 a,inAF3 b,inAF3 c){d=ALerpF3(a,b,c);return d;} + AF4 opALerpF4(outAF4 d,inAF4 a,inAF4 b,inAF4 c){d=ALerpF4(a,b,c);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opALerpOneF2(outAF2 d,inAF2 a,inAF2 b,AF1 c){d=ALerpF2(a,b,AF2_(c));return d;} + AF3 opALerpOneF3(outAF3 d,inAF3 a,inAF3 b,AF1 c){d=ALerpF3(a,b,AF3_(c));return d;} + AF4 opALerpOneF4(outAF4 d,inAF4 a,inAF4 b,AF1 c){d=ALerpF4(a,b,AF4_(c));return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMaxF2(outAF2 d,inAF2 a,inAF2 b){d=max(a,b);return d;} + AF3 opAMaxF3(outAF3 d,inAF3 a,inAF3 b){d=max(a,b);return d;} + AF4 opAMaxF4(outAF4 d,inAF4 a,inAF4 b){d=max(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMinF2(outAF2 d,inAF2 a,inAF2 b){d=min(a,b);return d;} + AF3 opAMinF3(outAF3 d,inAF3 a,inAF3 b){d=min(a,b);return d;} + AF4 opAMinF4(outAF4 d,inAF4 a,inAF4 b){d=min(a,b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMulF2(outAF2 d,inAF2 a,inAF2 b){d=a*b;return d;} + AF3 opAMulF3(outAF3 d,inAF3 a,inAF3 b){d=a*b;return d;} + AF4 opAMulF4(outAF4 d,inAF4 a,inAF4 b){d=a*b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAMulOneF2(outAF2 d,inAF2 a,AF1 b){d=a*AF2_(b);return d;} + AF3 opAMulOneF3(outAF3 d,inAF3 a,AF1 b){d=a*AF3_(b);return d;} + AF4 opAMulOneF4(outAF4 d,inAF4 a,AF1 b){d=a*AF4_(b);return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opANegF2(outAF2 d,inAF2 a){d=-a;return d;} + AF3 opANegF3(outAF3 d,inAF3 a){d=-a;return d;} + AF4 opANegF4(outAF4 d,inAF4 a){d=-a;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opARcpF2(outAF2 d,inAF2 a){d=ARcpF2(a);return d;} + AF3 opARcpF3(outAF3 d,inAF3 a){d=ARcpF3(a);return d;} + AF4 opARcpF4(outAF4 d,inAF4 a){d=ARcpF4(a);return d;} +#endif + + +#define FSR_RCAS_F 1 +AU4 con0; + +AF4 FsrRcasLoadF(ASU2 p) { return AF4(texelFetch(source, p, 0)); } +void FsrRcasInputF(inout AF1 r, inout AF1 g, inout AF1 b) {} + +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// AMD FidelityFX SUPER RESOLUTION [FSR 1] ::: SPATIAL SCALING & EXTRAS - v1.20210629 +// +// +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// FidelityFX Super Resolution Sample +// +// Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved. +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files(the "Software"), to deal +// in the Software without restriction, including without limitation the rights +// to use, copy, modify, merge, publish, distribute, sublicense, and / or sell +// copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions : +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN +// THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// ABOUT +// ===== +// FSR is a collection of algorithms relating to generating a higher resolution image. +// This specific header focuses on single-image non-temporal image scaling, and related tools. +// +// The core functions are EASU and RCAS: +// [EASU] Edge Adaptive Spatial Upsampling ....... 1x to 4x area range spatial scaling, clamped adaptive elliptical filter. +// [RCAS] Robust Contrast Adaptive Sharpening .... A non-scaling variation on CAS. +// RCAS needs to be applied after EASU as a separate pass. +// +// Optional utility functions are: +// [LFGA] Linear Film Grain Applicator ........... Tool to apply film grain after scaling. +// [SRTM] Simple Reversible Tone-Mapper .......... Linear HDR {0 to FP16_MAX} to {0 to 1} and back. +// [TEPD] Temporal Energy Preserving Dither ...... Temporally energy preserving dithered {0 to 1} linear to gamma 2.0 conversion. +// See each individual sub-section for inline documentation. +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// FUNCTION PERMUTATIONS +// ===================== +// *F() ..... Single item computation with 32-bit. +// *H() ..... Single item computation with 16-bit, with packing (aka two 16-bit ops in parallel) when possible. +// *Hx2() ... Processing two items in parallel with 16-bit, easier packing. +// Not all interfaces in this file have a *Hx2() form. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [EASU] EDGE ADAPTIVE SPATIAL UPSAMPLING +// +//------------------------------------------------------------------------------------------------------------------------------ +// EASU provides a high quality spatial-only scaling at relatively low cost. +// Meaning EASU is appropiate for laptops and other low-end GPUs. +// Quality from 1x to 4x area scaling is good. +//------------------------------------------------------------------------------------------------------------------------------ +// The scalar uses a modified fast approximation to the standard lanczos(size=2) kernel. +// EASU runs in a single pass, so it applies a directionally and anisotropically adaptive radial lanczos. +// This is also kept as simple as possible to have minimum runtime. +//------------------------------------------------------------------------------------------------------------------------------ +// The lanzcos filter has negative lobes, so by itself it will introduce ringing. +// To remove all ringing, the algorithm uses the nearest 2x2 input texels as a neighborhood, +// and limits output to the minimum and maximum of that neighborhood. +//------------------------------------------------------------------------------------------------------------------------------ +// Input image requirements: +// +// Color needs to be encoded as 3 channel[red, green, blue](e.g.XYZ not supported) +// Each channel needs to be in the range[0, 1] +// Any color primaries are supported +// Display / tonemapping curve needs to be as if presenting to sRGB display or similar(e.g.Gamma 2.0) +// There should be no banding in the input +// There should be no high amplitude noise in the input +// There should be no noise in the input that is not at input pixel granularity +// For performance purposes, use 32bpp formats +//------------------------------------------------------------------------------------------------------------------------------ +// Best to apply EASU at the end of the frame after tonemapping +// but before film grain or composite of the UI. +//------------------------------------------------------------------------------------------------------------------------------ +// Example of including this header for D3D HLSL : +// +// #define A_GPU 1 +// #define A_HLSL 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of including this header for Vulkan GLSL : +// +// #define A_GPU 1 +// #define A_GLSL 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of including this header for Vulkan HLSL : +// +// #define A_GPU 1 +// #define A_HLSL 1 +// #define A_HLSL_6_2 1 +// #define A_NO_16_BIT_CAST 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of declaring the required input callbacks for GLSL : +// The callbacks need to gather4 for each color channel using the specified texture coordinate 'p'. +// EASU uses gather4 to reduce position computation logic and for free Arrays of Structures to Structures of Arrays conversion. +// +// AH4 FsrEasuRH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,0));} +// AH4 FsrEasuGH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,1));} +// AH4 FsrEasuBH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,2));} +// ... +// The FsrEasuCon function needs to be called from the CPU or GPU to set up constants. +// The difference in viewport and input image size is there to support Dynamic Resolution Scaling. +// To use FsrEasuCon() on the CPU, define A_CPU before including ffx_a and ffx_fsr1. +// Including a GPU example here, the 'con0' through 'con3' values would be stored out to a constant buffer. +// AU4 con0,con1,con2,con3; +// FsrEasuCon(con0,con1,con2,con3, +// 1920.0,1080.0, // Viewport size (top left aligned) in the input image which is to be scaled. +// 3840.0,2160.0, // The size of the input image. +// 2560.0,1440.0); // The output resolution. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CONSTANT SETUP +//============================================================================================================================== +// Call to setup required constant values (works on CPU or GPU). +A_STATIC void FsrEasuCon( +outAU4 con0, +outAU4 con1, +outAU4 con2, +outAU4 con3, +// This the rendered image resolution being upscaled +AF1 inputViewportInPixelsX, +AF1 inputViewportInPixelsY, +// This is the resolution of the resource containing the input image (useful for dynamic resolution) +AF1 inputSizeInPixelsX, +AF1 inputSizeInPixelsY, +// This is the display resolution which the input image gets upscaled to +AF1 outputSizeInPixelsX, +AF1 outputSizeInPixelsY){ + // Output integer position to a pixel position in viewport. + con0[0]=AU1_AF1(inputViewportInPixelsX*ARcpF1(outputSizeInPixelsX)); + con0[1]=AU1_AF1(inputViewportInPixelsY*ARcpF1(outputSizeInPixelsY)); + con0[2]=AU1_AF1(AF1_(0.5)*inputViewportInPixelsX*ARcpF1(outputSizeInPixelsX)-AF1_(0.5)); + con0[3]=AU1_AF1(AF1_(0.5)*inputViewportInPixelsY*ARcpF1(outputSizeInPixelsY)-AF1_(0.5)); + // Viewport pixel position to normalized image space. + // This is used to get upper-left of 'F' tap. + con1[0]=AU1_AF1(ARcpF1(inputSizeInPixelsX)); + con1[1]=AU1_AF1(ARcpF1(inputSizeInPixelsY)); + // Centers of gather4, first offset from upper-left of 'F'. + // +---+---+ + // | | | + // +--(0)--+ + // | b | c | + // +---F---+---+---+ + // | e | f | g | h | + // +--(1)--+--(2)--+ + // | i | j | k | l | + // +---+---+---+---+ + // | n | o | + // +--(3)--+ + // | | | + // +---+---+ + con1[2]=AU1_AF1(AF1_( 1.0)*ARcpF1(inputSizeInPixelsX)); + con1[3]=AU1_AF1(AF1_(-1.0)*ARcpF1(inputSizeInPixelsY)); + // These are from (0) instead of 'F'. + con2[0]=AU1_AF1(AF1_(-1.0)*ARcpF1(inputSizeInPixelsX)); + con2[1]=AU1_AF1(AF1_( 2.0)*ARcpF1(inputSizeInPixelsY)); + con2[2]=AU1_AF1(AF1_( 1.0)*ARcpF1(inputSizeInPixelsX)); + con2[3]=AU1_AF1(AF1_( 2.0)*ARcpF1(inputSizeInPixelsY)); + con3[0]=AU1_AF1(AF1_( 0.0)*ARcpF1(inputSizeInPixelsX)); + con3[1]=AU1_AF1(AF1_( 4.0)*ARcpF1(inputSizeInPixelsY)); + con3[2]=con3[3]=0;} + +//If the an offset into the input image resource +A_STATIC void FsrEasuConOffset( + outAU4 con0, + outAU4 con1, + outAU4 con2, + outAU4 con3, + // This the rendered image resolution being upscaled + AF1 inputViewportInPixelsX, + AF1 inputViewportInPixelsY, + // This is the resolution of the resource containing the input image (useful for dynamic resolution) + AF1 inputSizeInPixelsX, + AF1 inputSizeInPixelsY, + // This is the display resolution which the input image gets upscaled to + AF1 outputSizeInPixelsX, + AF1 outputSizeInPixelsY, + // This is the input image offset into the resource containing it (useful for dynamic resolution) + AF1 inputOffsetInPixelsX, + AF1 inputOffsetInPixelsY) { + FsrEasuCon(con0, con1, con2, con3, inputViewportInPixelsX, inputViewportInPixelsY, inputSizeInPixelsX, inputSizeInPixelsY, outputSizeInPixelsX, outputSizeInPixelsY); + con0[2] = AU1_AF1(AF1_(0.5) * inputViewportInPixelsX * ARcpF1(outputSizeInPixelsX) - AF1_(0.5) + inputOffsetInPixelsX); + con0[3] = AU1_AF1(AF1_(0.5) * inputViewportInPixelsY * ARcpF1(outputSizeInPixelsY) - AF1_(0.5) + inputOffsetInPixelsY); +} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 32-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(FSR_EASU_F) + // Input callback prototypes, need to be implemented by calling shader + AF4 FsrEasuRF(AF2 p); + AF4 FsrEasuGF(AF2 p); + AF4 FsrEasuBF(AF2 p); +//------------------------------------------------------------------------------------------------------------------------------ + // Filtering for a given tap for the scalar. + void FsrEasuTapF( + inout AF3 aC, // Accumulated color, with negative lobe. + inout AF1 aW, // Accumulated weight. + AF2 off, // Pixel offset from resolve position to tap. + AF2 dir, // Gradient direction. + AF2 len, // Length. + AF1 lob, // Negative lobe strength. + AF1 clp, // Clipping point. + AF3 c){ // Tap color. + // Rotate offset by direction. + AF2 v; + v.x=(off.x*( dir.x))+(off.y*dir.y); + v.y=(off.x*(-dir.y))+(off.y*dir.x); + // Anisotropy. + v*=len; + // Compute distance^2. + AF1 d2=v.x*v.x+v.y*v.y; + // Limit to the window as at corner, 2 taps can easily be outside. + d2=min(d2,clp); + // Approximation of lancos2 without sin() or rcp(), or sqrt() to get x. + // (25/16 * (2/5 * x^2 - 1)^2 - (25/16 - 1)) * (1/4 * x^2 - 1)^2 + // |_______________________________________| |_______________| + // base window + // The general form of the 'base' is, + // (a*(b*x^2-1)^2-(a-1)) + // Where 'a=1/(2*b-b^2)' and 'b' moves around the negative lobe. + AF1 wB=AF1_(2.0/5.0)*d2+AF1_(-1.0); + AF1 wA=lob*d2+AF1_(-1.0); + wB*=wB; + wA*=wA; + wB=AF1_(25.0/16.0)*wB+AF1_(-(25.0/16.0-1.0)); + AF1 w=wB*wA; + // Do weighted average. + aC+=c*w;aW+=w;} +//------------------------------------------------------------------------------------------------------------------------------ + // Accumulate direction and length. + void FsrEasuSetF( + inout AF2 dir, + inout AF1 len, + AF2 pp, + AP1 biS,AP1 biT,AP1 biU,AP1 biV, + AF1 lA,AF1 lB,AF1 lC,AF1 lD,AF1 lE){ + // Compute bilinear weight, branches factor out as predicates are compiler time immediates. + // s t + // u v + AF1 w = AF1_(0.0); + if(biS)w=(AF1_(1.0)-pp.x)*(AF1_(1.0)-pp.y); + if(biT)w= pp.x *(AF1_(1.0)-pp.y); + if(biU)w=(AF1_(1.0)-pp.x)* pp.y ; + if(biV)w= pp.x * pp.y ; + // Direction is the '+' diff. + // a + // b c d + // e + // Then takes magnitude from abs average of both sides of 'c'. + // Length converts gradient reversal to 0, smoothly to non-reversal at 1, shaped, then adding horz and vert terms. + AF1 dc=lD-lC; + AF1 cb=lC-lB; + AF1 lenX=max(abs(dc),abs(cb)); + lenX=APrxLoRcpF1(lenX); + AF1 dirX=lD-lB; + dir.x+=dirX*w; + lenX=ASatF1(abs(dirX)*lenX); + lenX*=lenX; + len+=lenX*w; + // Repeat for the y axis. + AF1 ec=lE-lC; + AF1 ca=lC-lA; + AF1 lenY=max(abs(ec),abs(ca)); + lenY=APrxLoRcpF1(lenY); + AF1 dirY=lE-lA; + dir.y+=dirY*w; + lenY=ASatF1(abs(dirY)*lenY); + lenY*=lenY; + len+=lenY*w;} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrEasuF( + out AF3 pix, + AU2 ip, // Integer pixel position in output. + AU4 con0, // Constants generated by FsrEasuCon(). + AU4 con1, + AU4 con2, + AU4 con3){ +//------------------------------------------------------------------------------------------------------------------------------ + // Get position of 'f'. + AF2 pp=AF2(ip)*AF2_AU2(con0.xy)+AF2_AU2(con0.zw); + AF2 fp=floor(pp); + pp-=fp; +//------------------------------------------------------------------------------------------------------------------------------ + // 12-tap kernel. + // b c + // e f g h + // i j k l + // n o + // Gather 4 ordering. + // a b + // r g + // For packed FP16, need either {rg} or {ab} so using the following setup for gather in all versions, + // a b <- unused (z) + // r g + // a b a b + // r g r g + // a b + // r g <- unused (z) + // Allowing dead-code removal to remove the 'z's. + AF2 p0=fp*AF2_AU2(con1.xy)+AF2_AU2(con1.zw); + // These are from p0 to avoid pulling two constants on pre-Navi hardware. + AF2 p1=p0+AF2_AU2(con2.xy); + AF2 p2=p0+AF2_AU2(con2.zw); + AF2 p3=p0+AF2_AU2(con3.xy); + AF4 bczzR=FsrEasuRF(p0); + AF4 bczzG=FsrEasuGF(p0); + AF4 bczzB=FsrEasuBF(p0); + AF4 ijfeR=FsrEasuRF(p1); + AF4 ijfeG=FsrEasuGF(p1); + AF4 ijfeB=FsrEasuBF(p1); + AF4 klhgR=FsrEasuRF(p2); + AF4 klhgG=FsrEasuGF(p2); + AF4 klhgB=FsrEasuBF(p2); + AF4 zzonR=FsrEasuRF(p3); + AF4 zzonG=FsrEasuGF(p3); + AF4 zzonB=FsrEasuBF(p3); +//------------------------------------------------------------------------------------------------------------------------------ + // Simplest multi-channel approximate luma possible (luma times 2, in 2 FMA/MAD). + AF4 bczzL=bczzB*AF4_(0.5)+(bczzR*AF4_(0.5)+bczzG); + AF4 ijfeL=ijfeB*AF4_(0.5)+(ijfeR*AF4_(0.5)+ijfeG); + AF4 klhgL=klhgB*AF4_(0.5)+(klhgR*AF4_(0.5)+klhgG); + AF4 zzonL=zzonB*AF4_(0.5)+(zzonR*AF4_(0.5)+zzonG); + // Rename. + AF1 bL=bczzL.x; + AF1 cL=bczzL.y; + AF1 iL=ijfeL.x; + AF1 jL=ijfeL.y; + AF1 fL=ijfeL.z; + AF1 eL=ijfeL.w; + AF1 kL=klhgL.x; + AF1 lL=klhgL.y; + AF1 hL=klhgL.z; + AF1 gL=klhgL.w; + AF1 oL=zzonL.z; + AF1 nL=zzonL.w; + // Accumulate for bilinear interpolation. + AF2 dir=AF2_(0.0); + AF1 len=AF1_(0.0); + FsrEasuSetF(dir,len,pp,true, false,false,false,bL,eL,fL,gL,jL); + FsrEasuSetF(dir,len,pp,false,true ,false,false,cL,fL,gL,hL,kL); + FsrEasuSetF(dir,len,pp,false,false,true ,false,fL,iL,jL,kL,nL); + FsrEasuSetF(dir,len,pp,false,false,false,true ,gL,jL,kL,lL,oL); +//------------------------------------------------------------------------------------------------------------------------------ + // Normalize with approximation, and cleanup close to zero. + AF2 dir2=dir*dir; + AF1 dirR=dir2.x+dir2.y; + AP1 zro=dirR w = -m/(n+e+w+s) +// 1 == (w*(n+e+w+s)+m)/(4*w+1) -> w = (1-m)/(n+e+w+s-4*1) +// Then chooses the 'w' which results in no clipping, limits 'w', and multiplies by the 'sharp' amount. +// This solution above has issues with MSAA input as the steps along the gradient cause edge detection issues. +// So RCAS uses 4x the maximum and 4x the minimum (depending on equation)in place of the individual taps. +// As well as switching from 'm' to either the minimum or maximum (depending on side), to help in energy conservation. +// This stabilizes RCAS. +// RCAS does a simple highpass which is normalized against the local contrast then shaped, +// 0.25 +// 0.25 -1 0.25 +// 0.25 +// This is used as a noise detection filter, to reduce the effect of RCAS on grain, and focus on real edges. +// +// GLSL example for the required callbacks : +// +// AH4 FsrRcasLoadH(ASW2 p){return AH4(imageLoad(imgSrc,ASU2(p)));} +// void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b) +// { +// //do any simple input color conversions here or leave empty if none needed +// } +// +// FsrRcasCon need to be called from the CPU or GPU to set up constants. +// Including a GPU example here, the 'con' value would be stored out to a constant buffer. +// +// AU4 con; +// FsrRcasCon(con, +// 0.0); // The scale is {0.0 := maximum sharpness, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. +// --------------- +// RCAS sharpening supports a CAS-like pass-through alpha via, +// #define FSR_RCAS_PASSTHROUGH_ALPHA 1 +// RCAS also supports a define to enable a more expensive path to avoid some sharpening of noise. +// Would suggest it is better to apply film grain after RCAS sharpening (and after scaling) instead of using this define, +// #define FSR_RCAS_DENOISE 1 +//============================================================================================================================== +// This is set at the limit of providing unnatural results for sharpening. +#define FSR_RCAS_LIMIT (0.25-(1.0/16.0)) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CONSTANT SETUP +//============================================================================================================================== +// Call to setup required constant values (works on CPU or GPU). +A_STATIC void FsrRcasCon( +outAU4 con, +// The scale is {0.0 := maximum, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. +AF1 sharpness){ + // Transform from stops to linear value. + sharpness=AExp2F1(-sharpness); + varAF2(hSharp)=initAF2(sharpness,sharpness); + con[0]=AU1_AF1(sharpness); + con[1]=AU1_AH2_AF2(hSharp); + con[2]=0; + con[3]=0;} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 32-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(FSR_RCAS_F) + // Input callback prototypes that need to be implemented by calling shader + AF4 FsrRcasLoadF(ASU2 p); + void FsrRcasInputF(inout AF1 r,inout AF1 g,inout AF1 b); +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasF( + out AF1 pixR, // Output values, non-vector so port between RcasFilter() and RcasFilterH() is easy. + out AF1 pixG, + out AF1 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AF1 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // Algorithm uses minimal 3x3 pixel neighborhood. + // b + // d e f + // h + ASU2 sp=ASU2(ip); + AF3 b=FsrRcasLoadF(sp+ASU2( 0,-1)).rgb; + AF3 d=FsrRcasLoadF(sp+ASU2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AF4 ee=FsrRcasLoadF(sp); + AF3 e=ee.rgb;pixA=ee.a; + #else + AF3 e=FsrRcasLoadF(sp).rgb; + #endif + AF3 f=FsrRcasLoadF(sp+ASU2( 1, 0)).rgb; + AF3 h=FsrRcasLoadF(sp+ASU2( 0, 1)).rgb; + // Rename (32-bit) or regroup (16-bit). + AF1 bR=b.r; + AF1 bG=b.g; + AF1 bB=b.b; + AF1 dR=d.r; + AF1 dG=d.g; + AF1 dB=d.b; + AF1 eR=e.r; + AF1 eG=e.g; + AF1 eB=e.b; + AF1 fR=f.r; + AF1 fG=f.g; + AF1 fB=f.b; + AF1 hR=h.r; + AF1 hG=h.g; + AF1 hB=h.b; + // Run optional input transform. + FsrRcasInputF(bR,bG,bB); + FsrRcasInputF(dR,dG,dB); + FsrRcasInputF(eR,eG,eB); + FsrRcasInputF(fR,fG,fB); + FsrRcasInputF(hR,hG,hB); + // Luma times 2. + AF1 bL=bB*AF1_(0.5)+(bR*AF1_(0.5)+bG); + AF1 dL=dB*AF1_(0.5)+(dR*AF1_(0.5)+dG); + AF1 eL=eB*AF1_(0.5)+(eR*AF1_(0.5)+eG); + AF1 fL=fB*AF1_(0.5)+(fR*AF1_(0.5)+fG); + AF1 hL=hB*AF1_(0.5)+(hR*AF1_(0.5)+hG); + // Noise detection. + AF1 nz=AF1_(0.25)*bL+AF1_(0.25)*dL+AF1_(0.25)*fL+AF1_(0.25)*hL-eL; + nz=ASatF1(abs(nz)*APrxMedRcpF1(AMax3F1(AMax3F1(bL,dL,eL),fL,hL)-AMin3F1(AMin3F1(bL,dL,eL),fL,hL))); + nz=AF1_(-0.5)*nz+AF1_(1.0); + // Min and max of ring. + AF1 mn4R=min(AMin3F1(bR,dR,fR),hR); + AF1 mn4G=min(AMin3F1(bG,dG,fG),hG); + AF1 mn4B=min(AMin3F1(bB,dB,fB),hB); + AF1 mx4R=max(AMax3F1(bR,dR,fR),hR); + AF1 mx4G=max(AMax3F1(bG,dG,fG),hG); + AF1 mx4B=max(AMax3F1(bB,dB,fB),hB); + // Immediate constants for peak range. + AF2 peakC=AF2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AF1 hitMinR=min(mn4R,eR)*ARcpF1(AF1_(4.0)*mx4R); + AF1 hitMinG=min(mn4G,eG)*ARcpF1(AF1_(4.0)*mx4G); + AF1 hitMinB=min(mn4B,eB)*ARcpF1(AF1_(4.0)*mx4B); + AF1 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpF1(AF1_(4.0)*mn4R+peakC.y); + AF1 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpF1(AF1_(4.0)*mn4G+peakC.y); + AF1 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpF1(AF1_(4.0)*mn4B+peakC.y); + AF1 lobeR=max(-hitMinR,hitMaxR); + AF1 lobeG=max(-hitMinG,hitMaxG); + AF1 lobeB=max(-hitMinB,hitMaxB); + AF1 lobe=max(AF1_(-FSR_RCAS_LIMIT),min(AMax3F1(lobeR,lobeG,lobeB),AF1_(0.0)))*AF1_AU1(con.x); + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AF1 rcpL=APrxMedRcpF1(AF1_(4.0)*lobe+AF1_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL; + return;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 16-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_RCAS_H) + // Input callback prototypes that need to be implemented by calling shader + AH4 FsrRcasLoadH(ASW2 p); + void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b); +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasH( + out AH1 pixR, // Output values, non-vector so port between RcasFilter() and RcasFilterH() is easy. + out AH1 pixG, + out AH1 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AH1 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // Sharpening algorithm uses minimal 3x3 pixel neighborhood. + // b + // d e f + // h + ASW2 sp=ASW2(ip); + AH3 b=FsrRcasLoadH(sp+ASW2( 0,-1)).rgb; + AH3 d=FsrRcasLoadH(sp+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee=FsrRcasLoadH(sp); + AH3 e=ee.rgb;pixA=ee.a; + #else + AH3 e=FsrRcasLoadH(sp).rgb; + #endif + AH3 f=FsrRcasLoadH(sp+ASW2( 1, 0)).rgb; + AH3 h=FsrRcasLoadH(sp+ASW2( 0, 1)).rgb; + // Rename (32-bit) or regroup (16-bit). + AH1 bR=b.r; + AH1 bG=b.g; + AH1 bB=b.b; + AH1 dR=d.r; + AH1 dG=d.g; + AH1 dB=d.b; + AH1 eR=e.r; + AH1 eG=e.g; + AH1 eB=e.b; + AH1 fR=f.r; + AH1 fG=f.g; + AH1 fB=f.b; + AH1 hR=h.r; + AH1 hG=h.g; + AH1 hB=h.b; + // Run optional input transform. + FsrRcasInputH(bR,bG,bB); + FsrRcasInputH(dR,dG,dB); + FsrRcasInputH(eR,eG,eB); + FsrRcasInputH(fR,fG,fB); + FsrRcasInputH(hR,hG,hB); + // Luma times 2. + AH1 bL=bB*AH1_(0.5)+(bR*AH1_(0.5)+bG); + AH1 dL=dB*AH1_(0.5)+(dR*AH1_(0.5)+dG); + AH1 eL=eB*AH1_(0.5)+(eR*AH1_(0.5)+eG); + AH1 fL=fB*AH1_(0.5)+(fR*AH1_(0.5)+fG); + AH1 hL=hB*AH1_(0.5)+(hR*AH1_(0.5)+hG); + // Noise detection. + AH1 nz=AH1_(0.25)*bL+AH1_(0.25)*dL+AH1_(0.25)*fL+AH1_(0.25)*hL-eL; + nz=ASatH1(abs(nz)*APrxMedRcpH1(AMax3H1(AMax3H1(bL,dL,eL),fL,hL)-AMin3H1(AMin3H1(bL,dL,eL),fL,hL))); + nz=AH1_(-0.5)*nz+AH1_(1.0); + // Min and max of ring. + AH1 mn4R=min(AMin3H1(bR,dR,fR),hR); + AH1 mn4G=min(AMin3H1(bG,dG,fG),hG); + AH1 mn4B=min(AMin3H1(bB,dB,fB),hB); + AH1 mx4R=max(AMax3H1(bR,dR,fR),hR); + AH1 mx4G=max(AMax3H1(bG,dG,fG),hG); + AH1 mx4B=max(AMax3H1(bB,dB,fB),hB); + // Immediate constants for peak range. + AH2 peakC=AH2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AH1 hitMinR=min(mn4R,eR)*ARcpH1(AH1_(4.0)*mx4R); + AH1 hitMinG=min(mn4G,eG)*ARcpH1(AH1_(4.0)*mx4G); + AH1 hitMinB=min(mn4B,eB)*ARcpH1(AH1_(4.0)*mx4B); + AH1 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpH1(AH1_(4.0)*mn4R+peakC.y); + AH1 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpH1(AH1_(4.0)*mn4G+peakC.y); + AH1 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpH1(AH1_(4.0)*mn4B+peakC.y); + AH1 lobeR=max(-hitMinR,hitMaxR); + AH1 lobeG=max(-hitMinG,hitMaxG); + AH1 lobeB=max(-hitMinB,hitMaxB); + AH1 lobe=max(AH1_(-FSR_RCAS_LIMIT),min(AMax3H1(lobeR,lobeG,lobeB),AH1_(0.0)))*AH2_AU1(con.y).x; + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AH1 rcpL=APrxMedRcpH1(AH1_(4.0)*lobe+AH1_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PACKED 16-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_RCAS_HX2) + // Input callback prototypes that need to be implemented by the calling shader + AH4 FsrRcasLoadHx2(ASW2 p); + void FsrRcasInputHx2(inout AH2 r,inout AH2 g,inout AH2 b); +//------------------------------------------------------------------------------------------------------------------------------ + // Can be used to convert from packed Structures of Arrays to Arrays of Structures for store. + void FsrRcasDepackHx2(out AH4 pix0,out AH4 pix1,AH2 pixR,AH2 pixG,AH2 pixB){ + #ifdef A_HLSL + // Invoke a slower path for DX only, since it won't allow uninitialized values. + pix0.a=pix1.a=0.0; + #endif + pix0.rgb=AH3(pixR.x,pixG.x,pixB.x); + pix1.rgb=AH3(pixR.y,pixG.y,pixB.y);} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasHx2( + // Output values are for 2 8x8 tiles in a 16x8 region. + // pix.x = left 8x8 tile + // pix.y = right 8x8 tile + // This enables later processing to easily be packed as well. + out AH2 pixR, + out AH2 pixG, + out AH2 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AH2 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // No scaling algorithm uses minimal 3x3 pixel neighborhood. + ASW2 sp0=ASW2(ip); + AH3 b0=FsrRcasLoadHx2(sp0+ASW2( 0,-1)).rgb; + AH3 d0=FsrRcasLoadHx2(sp0+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee0=FsrRcasLoadHx2(sp0); + AH3 e0=ee0.rgb;pixA.r=ee0.a; + #else + AH3 e0=FsrRcasLoadHx2(sp0).rgb; + #endif + AH3 f0=FsrRcasLoadHx2(sp0+ASW2( 1, 0)).rgb; + AH3 h0=FsrRcasLoadHx2(sp0+ASW2( 0, 1)).rgb; + ASW2 sp1=sp0+ASW2(8,0); + AH3 b1=FsrRcasLoadHx2(sp1+ASW2( 0,-1)).rgb; + AH3 d1=FsrRcasLoadHx2(sp1+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee1=FsrRcasLoadHx2(sp1); + AH3 e1=ee1.rgb;pixA.g=ee1.a; + #else + AH3 e1=FsrRcasLoadHx2(sp1).rgb; + #endif + AH3 f1=FsrRcasLoadHx2(sp1+ASW2( 1, 0)).rgb; + AH3 h1=FsrRcasLoadHx2(sp1+ASW2( 0, 1)).rgb; + // Arrays of Structures to Structures of Arrays conversion. + AH2 bR=AH2(b0.r,b1.r); + AH2 bG=AH2(b0.g,b1.g); + AH2 bB=AH2(b0.b,b1.b); + AH2 dR=AH2(d0.r,d1.r); + AH2 dG=AH2(d0.g,d1.g); + AH2 dB=AH2(d0.b,d1.b); + AH2 eR=AH2(e0.r,e1.r); + AH2 eG=AH2(e0.g,e1.g); + AH2 eB=AH2(e0.b,e1.b); + AH2 fR=AH2(f0.r,f1.r); + AH2 fG=AH2(f0.g,f1.g); + AH2 fB=AH2(f0.b,f1.b); + AH2 hR=AH2(h0.r,h1.r); + AH2 hG=AH2(h0.g,h1.g); + AH2 hB=AH2(h0.b,h1.b); + // Run optional input transform. + FsrRcasInputHx2(bR,bG,bB); + FsrRcasInputHx2(dR,dG,dB); + FsrRcasInputHx2(eR,eG,eB); + FsrRcasInputHx2(fR,fG,fB); + FsrRcasInputHx2(hR,hG,hB); + // Luma times 2. + AH2 bL=bB*AH2_(0.5)+(bR*AH2_(0.5)+bG); + AH2 dL=dB*AH2_(0.5)+(dR*AH2_(0.5)+dG); + AH2 eL=eB*AH2_(0.5)+(eR*AH2_(0.5)+eG); + AH2 fL=fB*AH2_(0.5)+(fR*AH2_(0.5)+fG); + AH2 hL=hB*AH2_(0.5)+(hR*AH2_(0.5)+hG); + // Noise detection. + AH2 nz=AH2_(0.25)*bL+AH2_(0.25)*dL+AH2_(0.25)*fL+AH2_(0.25)*hL-eL; + nz=ASatH2(abs(nz)*APrxMedRcpH2(AMax3H2(AMax3H2(bL,dL,eL),fL,hL)-AMin3H2(AMin3H2(bL,dL,eL),fL,hL))); + nz=AH2_(-0.5)*nz+AH2_(1.0); + // Min and max of ring. + AH2 mn4R=min(AMin3H2(bR,dR,fR),hR); + AH2 mn4G=min(AMin3H2(bG,dG,fG),hG); + AH2 mn4B=min(AMin3H2(bB,dB,fB),hB); + AH2 mx4R=max(AMax3H2(bR,dR,fR),hR); + AH2 mx4G=max(AMax3H2(bG,dG,fG),hG); + AH2 mx4B=max(AMax3H2(bB,dB,fB),hB); + // Immediate constants for peak range. + AH2 peakC=AH2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AH2 hitMinR=min(mn4R,eR)*ARcpH2(AH2_(4.0)*mx4R); + AH2 hitMinG=min(mn4G,eG)*ARcpH2(AH2_(4.0)*mx4G); + AH2 hitMinB=min(mn4B,eB)*ARcpH2(AH2_(4.0)*mx4B); + AH2 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpH2(AH2_(4.0)*mn4R+peakC.y); + AH2 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpH2(AH2_(4.0)*mn4G+peakC.y); + AH2 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpH2(AH2_(4.0)*mn4B+peakC.y); + AH2 lobeR=max(-hitMinR,hitMaxR); + AH2 lobeG=max(-hitMinG,hitMaxG); + AH2 lobeB=max(-hitMinB,hitMaxB); + AH2 lobe=max(AH2_(-FSR_RCAS_LIMIT),min(AMax3H2(lobeR,lobeG,lobeB),AH2_(0.0)))*AH2_(AH2_AU1(con.y).x); + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AH2 rcpL=APrxMedRcpH2(AH2_(4.0)*lobe+AH2_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [LFGA] LINEAR FILM GRAIN APPLICATOR +// +//------------------------------------------------------------------------------------------------------------------------------ +// Adding output-resolution film grain after scaling is a good way to mask both rendering and scaling artifacts. +// Suggest using tiled blue noise as film grain input, with peak noise frequency set for a specific look and feel. +// The 'Lfga*()' functions provide a convenient way to introduce grain. +// These functions limit grain based on distance to signal limits. +// This is done so that the grain is temporally energy preserving, and thus won't modify image tonality. +// Grain application should be done in a linear colorspace. +// The grain should be temporally changing, but have a temporal sum per pixel that adds to zero (non-biased). +//------------------------------------------------------------------------------------------------------------------------------ +// Usage, +// FsrLfga*( +// color, // In/out linear colorspace color {0 to 1} ranged. +// grain, // Per pixel grain texture value {-0.5 to 0.5} ranged, input is 3-channel to support colored grain. +// amount); // Amount of grain (0 to 1} ranged. +//------------------------------------------------------------------------------------------------------------------------------ +// Example if grain texture is monochrome: 'FsrLfgaF(color,AF3_(grain),amount)' +//============================================================================================================================== +#if defined(A_GPU) + // Maximum grain is the minimum distance to the signal limit. + void FsrLfgaF(inout AF3 c,AF3 t,AF1 a){c+=(t*AF3_(a))*min(AF3_(1.0)-c,c);} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + // Half precision version (slower). + void FsrLfgaH(inout AH3 c,AH3 t,AH1 a){c+=(t*AH3_(a))*min(AH3_(1.0)-c,c);} +//------------------------------------------------------------------------------------------------------------------------------ + // Packed half precision version (faster). + void FsrLfgaHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 tR,AH2 tG,AH2 tB,AH1 a){ + cR+=(tR*AH2_(a))*min(AH2_(1.0)-cR,cR);cG+=(tG*AH2_(a))*min(AH2_(1.0)-cG,cG);cB+=(tB*AH2_(a))*min(AH2_(1.0)-cB,cB);} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [SRTM] SIMPLE REVERSIBLE TONE-MAPPER +// +//------------------------------------------------------------------------------------------------------------------------------ +// This provides a way to take linear HDR color {0 to FP16_MAX} and convert it into a temporary {0 to 1} ranged post-tonemapped linear. +// The tonemapper preserves RGB ratio, which helps maintain HDR color bleed during filtering. +//------------------------------------------------------------------------------------------------------------------------------ +// Reversible tonemapper usage, +// FsrSrtm*(color); // {0 to FP16_MAX} converted to {0 to 1}. +// FsrSrtmInv*(color); // {0 to 1} converted into {0 to 32768, output peak safe for FP16}. +//============================================================================================================================== +#if defined(A_GPU) + void FsrSrtmF(inout AF3 c){c*=AF3_(ARcpF1(AMax3F1(c.r,c.g,c.b)+AF1_(1.0)));} + // The extra max solves the c=1.0 case (which is a /0). + void FsrSrtmInvF(inout AF3 c){c*=AF3_(ARcpF1(max(AF1_(1.0/32768.0),AF1_(1.0)-AMax3F1(c.r,c.g,c.b))));} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + void FsrSrtmH(inout AH3 c){c*=AH3_(ARcpH1(AMax3H1(c.r,c.g,c.b)+AH1_(1.0)));} + void FsrSrtmInvH(inout AH3 c){c*=AH3_(ARcpH1(max(AH1_(1.0/32768.0),AH1_(1.0)-AMax3H1(c.r,c.g,c.b))));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrSrtmHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB){ + AH2 rcp=ARcpH2(AMax3H2(cR,cG,cB)+AH2_(1.0));cR*=rcp;cG*=rcp;cB*=rcp;} + void FsrSrtmInvHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB){ + AH2 rcp=ARcpH2(max(AH2_(1.0/32768.0),AH2_(1.0)-AMax3H2(cR,cG,cB)));cR*=rcp;cG*=rcp;cB*=rcp;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [TEPD] TEMPORAL ENERGY PRESERVING DITHER +// +//------------------------------------------------------------------------------------------------------------------------------ +// Temporally energy preserving dithered {0 to 1} linear to gamma 2.0 conversion. +// Gamma 2.0 is used so that the conversion back to linear is just to square the color. +// The conversion comes in 8-bit and 10-bit modes, designed for output to 8-bit UNORM or 10:10:10:2 respectively. +// Given good non-biased temporal blue noise as dither input, +// the output dither will temporally conserve energy. +// This is done by choosing the linear nearest step point instead of perceptual nearest. +// See code below for details. +//------------------------------------------------------------------------------------------------------------------------------ +// DX SPEC RULES FOR FLOAT->UNORM 8-BIT CONVERSION +// =============================================== +// - Output is 'uint(floor(saturate(n)*255.0+0.5))'. +// - Thus rounding is to nearest. +// - NaN gets converted to zero. +// - INF is clamped to {0.0 to 1.0}. +//============================================================================================================================== +#if defined(A_GPU) + // Hand tuned integer position to dither value, with more values than simple checkerboard. + // Only 32-bit has enough precision for this compddation. + // Output is {0 to <1}. + AF1 FsrTepdDitF(AU2 p,AU1 f){ + AF1 x=AF1_(p.x+f); + AF1 y=AF1_(p.y); + // The 1.61803 golden ratio. + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + // Number designed to provide a good visual pattern. + AF1 b=AF1_(1.0/3.69); + x=x*a+(y*b); + return AFractF1(x);} +//------------------------------------------------------------------------------------------------------------------------------ + // This version is 8-bit gamma 2.0. + // The 'c' input is {0 to 1}. + // Output is {0 to 1} ready for image store. + void FsrTepdC8F(inout AF3 c,AF1 dit){ + AF3 n=sqrt(c); + n=floor(n*AF3_(255.0))*AF3_(1.0/255.0); + AF3 a=n*n; + AF3 b=n+AF3_(1.0/255.0);b=b*b; + // Ratio of 'a' to 'b' required to produce 'c'. + // APrxLoRcpF1() won't work here (at least for very high dynamic ranges). + // APrxMedRcpF1() is an IADD,FMA,MUL. + AF3 r=(c-b)*APrxMedRcpF3(a-b); + // Use the ratio as a cutoff to choose 'a' or 'b'. + // AGtZeroF1() is a MUL. + c=ASatF3(n+AGtZeroF3(AF3_(dit)-r)*AF3_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + // This version is 10-bit gamma 2.0. + // The 'c' input is {0 to 1}. + // Output is {0 to 1} ready for image store. + void FsrTepdC10F(inout AF3 c,AF1 dit){ + AF3 n=sqrt(c); + n=floor(n*AF3_(1023.0))*AF3_(1.0/1023.0); + AF3 a=n*n; + AF3 b=n+AF3_(1.0/1023.0);b=b*b; + AF3 r=(c-b)*APrxMedRcpF3(a-b); + c=ASatF3(n+AGtZeroF3(AF3_(dit)-r)*AF3_(1.0/1023.0));} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + AH1 FsrTepdDitH(AU2 p,AU1 f){ + AF1 x=AF1_(p.x+f); + AF1 y=AF1_(p.y); + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + AF1 b=AF1_(1.0/3.69); + x=x*a+(y*b); + return AH1(AFractF1(x));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC8H(inout AH3 c,AH1 dit){ + AH3 n=sqrt(c); + n=floor(n*AH3_(255.0))*AH3_(1.0/255.0); + AH3 a=n*n; + AH3 b=n+AH3_(1.0/255.0);b=b*b; + AH3 r=(c-b)*APrxMedRcpH3(a-b); + c=ASatH3(n+AGtZeroH3(AH3_(dit)-r)*AH3_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC10H(inout AH3 c,AH1 dit){ + AH3 n=sqrt(c); + n=floor(n*AH3_(1023.0))*AH3_(1.0/1023.0); + AH3 a=n*n; + AH3 b=n+AH3_(1.0/1023.0);b=b*b; + AH3 r=(c-b)*APrxMedRcpH3(a-b); + c=ASatH3(n+AGtZeroH3(AH3_(dit)-r)*AH3_(1.0/1023.0));} +//============================================================================================================================== + // This computes dither for positions 'p' and 'p+{8,0}'. + AH2 FsrTepdDitHx2(AU2 p,AU1 f){ + AF2 x; + x.x=AF1_(p.x+f); + x.y=x.x+AF1_(8.0); + AF1 y=AF1_(p.y); + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + AF1 b=AF1_(1.0/3.69); + x=x*AF2_(a)+AF2_(y*b); + return AH2(AFractF2(x));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC8Hx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 dit){ + AH2 nR=sqrt(cR); + AH2 nG=sqrt(cG); + AH2 nB=sqrt(cB); + nR=floor(nR*AH2_(255.0))*AH2_(1.0/255.0); + nG=floor(nG*AH2_(255.0))*AH2_(1.0/255.0); + nB=floor(nB*AH2_(255.0))*AH2_(1.0/255.0); + AH2 aR=nR*nR; + AH2 aG=nG*nG; + AH2 aB=nB*nB; + AH2 bR=nR+AH2_(1.0/255.0);bR=bR*bR; + AH2 bG=nG+AH2_(1.0/255.0);bG=bG*bG; + AH2 bB=nB+AH2_(1.0/255.0);bB=bB*bB; + AH2 rR=(cR-bR)*APrxMedRcpH2(aR-bR); + AH2 rG=(cG-bG)*APrxMedRcpH2(aG-bG); + AH2 rB=(cB-bB)*APrxMedRcpH2(aB-bB); + cR=ASatH2(nR+AGtZeroH2(dit-rR)*AH2_(1.0/255.0)); + cG=ASatH2(nG+AGtZeroH2(dit-rG)*AH2_(1.0/255.0)); + cB=ASatH2(nB+AGtZeroH2(dit-rB)*AH2_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC10Hx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 dit){ + AH2 nR=sqrt(cR); + AH2 nG=sqrt(cG); + AH2 nB=sqrt(cB); + nR=floor(nR*AH2_(1023.0))*AH2_(1.0/1023.0); + nG=floor(nG*AH2_(1023.0))*AH2_(1.0/1023.0); + nB=floor(nB*AH2_(1023.0))*AH2_(1.0/1023.0); + AH2 aR=nR*nR; + AH2 aG=nG*nG; + AH2 aB=nB*nB; + AH2 bR=nR+AH2_(1.0/1023.0);bR=bR*bR; + AH2 bG=nG+AH2_(1.0/1023.0);bG=bG*bG; + AH2 bB=nB+AH2_(1.0/1023.0);bB=bB*bB; + AH2 rR=(cR-bR)*APrxMedRcpH2(aR-bR); + AH2 rG=(cG-bG)*APrxMedRcpH2(aG-bG); + AH2 rB=(cB-bB)*APrxMedRcpH2(aB-bB); + cR=ASatH2(nR+AGtZeroH2(dit-rR)*AH2_(1.0/1023.0)); + cG=ASatH2(nG+AGtZeroH2(dit-rG)*AH2_(1.0/1023.0)); + cB=ASatH2(nB+AGtZeroH2(dit-rB)*AH2_(1.0/1023.0));} +#endif + + +void CurrFilter(AU2 pos) +{ + AF3 c; + FsrRcasF(c.r, c.g, c.b, pos, con0); + imageStore(imgOutput, ASU2(pos), AF4(c, 1)); +} + +void main() { + FsrRcasCon(con0, sharpening_data); + + AU2 gxy = ARmp8x8(gl_LocalInvocationID.x) + AU2(gl_WorkGroupID.x << 4u, gl_WorkGroupID.y << 4u); + CurrFilter(gxy); + gxy.x += 8u; + CurrFilter(gxy); + gxy.y += 8u; + CurrFilter(gxy); + gxy.x -= 8u; + CurrFilter(gxy); +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrSharpening.spv b/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrSharpening.spv new file mode 100644 index 00000000..b2e30e1f Binary files /dev/null and b/Ryujinx.Graphics.Vulkan/Effects/Shaders/FsrSharpening.spv differ diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/Fxaa.glsl b/Ryujinx.Graphics.Vulkan/Effects/Shaders/Fxaa.glsl new file mode 100644 index 00000000..f197c64c --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/Shaders/Fxaa.glsl @@ -0,0 +1,1177 @@ +/*============================================================================ + + + NVIDIA FXAA 3.11 by TIMOTHY LOTTES + + +------------------------------------------------------------------------------ +COPYRIGHT (C) 2010, 2011 NVIDIA CORPORATION. ALL RIGHTS RESERVED. +------------------------------------------------------------------------------ +TO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE LAW, THIS SOFTWARE IS PROVIDED +*AS IS* AND NVIDIA AND ITS SUPPLIERS DISCLAIM ALL WARRANTIES, EITHER EXPRESS +OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, IMPLIED WARRANTIES OF +MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. IN NO EVENT SHALL NVIDIA +OR ITS SUPPLIERS BE LIABLE FOR ANY SPECIAL, INCIDENTAL, INDIRECT, OR +CONSEQUENTIAL DAMAGES WHATSOEVER (INCLUDING, WITHOUT LIMITATION, DAMAGES FOR +LOSS OF BUSINESS PROFITS, BUSINESS INTERRUPTION, LOSS OF BUSINESS INFORMATION, +OR ANY OTHER PECUNIARY LOSS) ARISING OUT OF THE USE OF OR INABILITY TO USE +THIS SOFTWARE, EVEN IF NVIDIA HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH +DAMAGES. + +------------------------------------------------------------------------------ + INTEGRATION CHECKLIST +------------------------------------------------------------------------------ +(1.) +In the shader source, setup defines for the desired configuration. +When providing multiple shaders (for different presets), +simply setup the defines differently in multiple files. +Example, + + #define FXAA_PC 1 + #define FXAA_HLSL_5 1 + #define FXAA_QUALITY_PRESET 12 + +Or, + + #define FXAA_360 1 + +Or, + + #define FXAA_PS3 1 + +Etc. + +(2.) +Then include this file, + + #include "Fxaa3_11.h" + +(3.) +Then call the FXAA pixel shader from within your desired shader. +Look at the FXAA Quality FxaaPixelShader() for docs on inputs. +As for FXAA 3.11 all inputs for all shaders are the same +to enable easy porting between platforms. + + return FxaaPixelShader(...); + +(4.) +Insure pass prior to FXAA outputs RGBL (see next section). +Or use, + + #define FXAA_GREEN_AS_LUMA 1 + +(5.) +Setup engine to provide the following constants +which are used in the FxaaPixelShader() inputs, + + FxaaFloat2 fxaaQualityRcpFrame, + FxaaFloat4 fxaaConsoleRcpFrameOpt, + FxaaFloat4 fxaaConsoleRcpFrameOpt2, + FxaaFloat4 fxaaConsole360RcpFrameOpt2, + FxaaFloat fxaaQualitySubpix, + FxaaFloat fxaaQualityEdgeThreshold, + FxaaFloat fxaaQualityEdgeThresholdMin, + FxaaFloat fxaaConsoleEdgeSharpness, + FxaaFloat fxaaConsoleEdgeThreshold, + FxaaFloat fxaaConsoleEdgeThresholdMin, + FxaaFloat4 fxaaConsole360ConstDir + +Look at the FXAA Quality FxaaPixelShader() for docs on inputs. + +(6.) +Have FXAA vertex shader run as a full screen triangle, +and output "pos" and "fxaaConsolePosPos" +such that inputs in the pixel shader provide, + + // {xy} = center of pixel + FxaaFloat2 pos, + + // {xy_} = upper left of pixel + // {_zw} = lower right of pixel + FxaaFloat4 fxaaConsolePosPos, + +(7.) +Insure the texture sampler(s) used by FXAA are set to bilinear filtering. + + +------------------------------------------------------------------------------ + INTEGRATION - RGBL AND COLORSPACE +------------------------------------------------------------------------------ +FXAA3 requires RGBL as input unless the following is set, + + #define FXAA_GREEN_AS_LUMA 1 + +In which case the engine uses green in place of luma, +and requires RGB input is in a non-linear colorspace. + +RGB should be LDR (low dynamic range). +Specifically do FXAA after tonemapping. + +RGB data as returned by a texture fetch can be non-linear, +or linear when FXAA_GREEN_AS_LUMA is not set. +Note an "sRGB format" texture counts as linear, +because the result of a texture fetch is linear data. +Regular "RGBA8" textures in the sRGB colorspace are non-linear. + +If FXAA_GREEN_AS_LUMA is not set, +luma must be stored in the alpha channel prior to running FXAA. +This luma should be in a perceptual space (could be gamma 2.0). +Example pass before FXAA where output is gamma 2.0 encoded, + + color.rgb = ToneMap(color.rgb); // linear color output + color.rgb = sqrt(color.rgb); // gamma 2.0 color output + return color; + +To use FXAA, + + color.rgb = ToneMap(color.rgb); // linear color output + color.rgb = sqrt(color.rgb); // gamma 2.0 color output + color.a = dot(color.rgb, FxaaFloat3(0.299, 0.587, 0.114)); // compute luma + return color; + +Another example where output is linear encoded, +say for instance writing to an sRGB formated render target, +where the render target does the conversion back to sRGB after blending, + + color.rgb = ToneMap(color.rgb); // linear color output + return color; + +To use FXAA, + + color.rgb = ToneMap(color.rgb); // linear color output + color.a = sqrt(dot(color.rgb, FxaaFloat3(0.299, 0.587, 0.114))); // compute luma + return color; + +Getting luma correct is required for the algorithm to work correctly. + + +------------------------------------------------------------------------------ + BEING LINEARLY CORRECT? +------------------------------------------------------------------------------ +Applying FXAA to a framebuffer with linear RGB color will look worse. +This is very counter intuitive, but happends to be true in this case. +The reason is because dithering artifacts will be more visiable +in a linear colorspace. + + +------------------------------------------------------------------------------ + COMPLEX INTEGRATION +------------------------------------------------------------------------------ +Q. What if the engine is blending into RGB before wanting to run FXAA? + +A. In the last opaque pass prior to FXAA, + have the pass write out luma into alpha. + Then blend into RGB only. + FXAA should be able to run ok + assuming the blending pass did not any add aliasing. + This should be the common case for particles and common blending passes. + +A. Or use FXAA_GREEN_AS_LUMA. + +============================================================================*/ + +#version 430 core + +layout(local_size_x = 16, local_size_y = 16) in; +layout(rgba8, binding = 0, set = 3) uniform image2D imgOutput; + +layout(binding = 1, set = 2) uniform sampler2D inputImage; +layout(binding = 2) uniform invResolution +{ + vec2 invResolution_data; +}; + +#define FXAA_QUALITY_PRESET 12 +#define FXAA_GREEN_AS_LUMA 1 +#define FXAA_PC 1 +#define FXAA_GLSL_130 1 + + +/*============================================================================ + + INTEGRATION KNOBS + +/*==========================================================================*/ +#ifndef FXAA_PC + // + // FXAA Quality + // The high quality PC algorithm. + // + #define FXAA_PC 0 +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_GLSL_120 + #define FXAA_GLSL_120 0 +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_GLSL_130 + #define FXAA_GLSL_130 0 +#endif +/*==========================================================================*/ +#ifndef FXAA_GREEN_AS_LUMA + // + // For those using non-linear color, + // and either not able to get luma in alpha, or not wanting to, + // this enables FXAA to run using green as a proxy for luma. + // So with this enabled, no need to pack luma in alpha. + // + // This will turn off AA on anything which lacks some amount of green. + // Pure red and blue or combination of only R and B, will get no AA. + // + // Might want to lower the settings for both, + // fxaaConsoleEdgeThresholdMin + // fxaaQualityEdgeThresholdMin + // In order to insure AA does not get turned off on colors + // which contain a minor amount of green. + // + // 1 = On. + // 0 = Off. + // + #define FXAA_GREEN_AS_LUMA 0 +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_EARLY_EXIT + // + // Controls algorithm's early exit path. + // On PS3 turning this ON adds 2 cycles to the shader. + // On 360 turning this OFF adds 10ths of a millisecond to the shader. + // Turning this off on console will result in a more blurry image. + // So this defaults to on. + // + // 1 = On. + // 0 = Off. + // + #define FXAA_EARLY_EXIT 1 +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_DISCARD + // + // Only valid for PC OpenGL currently. + // Probably will not work when FXAA_GREEN_AS_LUMA = 1. + // + // 1 = Use discard on pixels which don't need AA. + // For APIs which enable concurrent TEX+ROP from same surface. + // 0 = Return unchanged color on pixels which don't need AA. + // + #define FXAA_DISCARD 0 +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_FAST_PIXEL_OFFSET + // + // Used for GLSL 120 only. + // + // 1 = GL API supports fast pixel offsets + // 0 = do not use fast pixel offsets + // + #ifdef GL_EXT_gpu_shader4 + #define FXAA_FAST_PIXEL_OFFSET 1 + #endif + #ifdef GL_NV_gpu_shader5 + #define FXAA_FAST_PIXEL_OFFSET 1 + #endif + #ifdef GL_ARB_gpu_shader5 + #define FXAA_FAST_PIXEL_OFFSET 1 + #endif + #ifndef FXAA_FAST_PIXEL_OFFSET + #define FXAA_FAST_PIXEL_OFFSET 0 + #endif +#endif +/*--------------------------------------------------------------------------*/ +#ifndef FXAA_GATHER4_ALPHA + // + // 1 = API supports gather4 on alpha channel. + // 0 = API does not support gather4 on alpha channel. + // + #if (FXAA_HLSL_5 == 1) + #define FXAA_GATHER4_ALPHA 1 + #endif + #ifdef GL_ARB_gpu_shader5 + #define FXAA_GATHER4_ALPHA 1 + #endif + #ifdef GL_NV_gpu_shader5 + #define FXAA_GATHER4_ALPHA 1 + #endif + #ifndef FXAA_GATHER4_ALPHA + #define FXAA_GATHER4_ALPHA 0 + #endif +#endif + +/*============================================================================ + FXAA QUALITY - TUNING KNOBS +------------------------------------------------------------------------------ +NOTE the other tuning knobs are now in the shader function inputs! +============================================================================*/ +#ifndef FXAA_QUALITY_PRESET + // + // Choose the quality preset. + // This needs to be compiled into the shader as it effects code. + // Best option to include multiple presets is to + // in each shader define the preset, then include this file. + // + // OPTIONS + // ----------------------------------------------------------------------- + // 10 to 15 - default medium dither (10=fastest, 15=highest quality) + // 20 to 29 - less dither, more expensive (20=fastest, 29=highest quality) + // 39 - no dither, very expensive + // + // NOTES + // ----------------------------------------------------------------------- + // 12 = slightly faster then FXAA 3.9 and higher edge quality (default) + // 13 = about same speed as FXAA 3.9 and better than 12 + // 23 = closest to FXAA 3.9 visually and performance wise + // _ = the lowest digit is directly related to performance + // _ = the highest digit is directly related to style + // + #define FXAA_QUALITY_PRESET 12 +#endif + + +/*============================================================================ + + FXAA QUALITY - PRESETS + +============================================================================*/ + +/*============================================================================ + FXAA QUALITY - MEDIUM DITHER PRESETS +============================================================================*/ +#if (FXAA_QUALITY_PRESET == 10) + #define FXAA_QUALITY_PS 3 + #define FXAA_QUALITY_P0 1.5 + #define FXAA_QUALITY_P1 3.0 + #define FXAA_QUALITY_P2 12.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 11) + #define FXAA_QUALITY_PS 4 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 3.0 + #define FXAA_QUALITY_P3 12.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 12) + #define FXAA_QUALITY_PS 5 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 4.0 + #define FXAA_QUALITY_P4 12.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 13) + #define FXAA_QUALITY_PS 6 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 4.0 + #define FXAA_QUALITY_P5 12.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 14) + #define FXAA_QUALITY_PS 7 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 2.0 + #define FXAA_QUALITY_P5 4.0 + #define FXAA_QUALITY_P6 12.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 15) + #define FXAA_QUALITY_PS 8 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 2.0 + #define FXAA_QUALITY_P5 2.0 + #define FXAA_QUALITY_P6 4.0 + #define FXAA_QUALITY_P7 12.0 +#endif + +/*============================================================================ + FXAA QUALITY - LOW DITHER PRESETS +============================================================================*/ +#if (FXAA_QUALITY_PRESET == 20) + #define FXAA_QUALITY_PS 3 + #define FXAA_QUALITY_P0 1.5 + #define FXAA_QUALITY_P1 2.0 + #define FXAA_QUALITY_P2 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 21) + #define FXAA_QUALITY_PS 4 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 22) + #define FXAA_QUALITY_PS 5 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 23) + #define FXAA_QUALITY_PS 6 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 2.0 + #define FXAA_QUALITY_P5 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 24) + #define FXAA_QUALITY_PS 7 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 2.0 + #define FXAA_QUALITY_P5 3.0 + #define FXAA_QUALITY_P6 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 25) + #define FXAA_QUALITY_PS 8 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 2.0 + #define FXAA_QUALITY_P5 2.0 + #define FXAA_QUALITY_P6 4.0 + #define FXAA_QUALITY_P7 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 26) + #define FXAA_QUALITY_PS 9 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 2.0 + #define FXAA_QUALITY_P5 2.0 + #define FXAA_QUALITY_P6 2.0 + #define FXAA_QUALITY_P7 4.0 + #define FXAA_QUALITY_P8 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 27) + #define FXAA_QUALITY_PS 10 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 2.0 + #define FXAA_QUALITY_P5 2.0 + #define FXAA_QUALITY_P6 2.0 + #define FXAA_QUALITY_P7 2.0 + #define FXAA_QUALITY_P8 4.0 + #define FXAA_QUALITY_P9 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 28) + #define FXAA_QUALITY_PS 11 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 2.0 + #define FXAA_QUALITY_P5 2.0 + #define FXAA_QUALITY_P6 2.0 + #define FXAA_QUALITY_P7 2.0 + #define FXAA_QUALITY_P8 2.0 + #define FXAA_QUALITY_P9 4.0 + #define FXAA_QUALITY_P10 8.0 +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_QUALITY_PRESET == 29) + #define FXAA_QUALITY_PS 12 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.5 + #define FXAA_QUALITY_P2 2.0 + #define FXAA_QUALITY_P3 2.0 + #define FXAA_QUALITY_P4 2.0 + #define FXAA_QUALITY_P5 2.0 + #define FXAA_QUALITY_P6 2.0 + #define FXAA_QUALITY_P7 2.0 + #define FXAA_QUALITY_P8 2.0 + #define FXAA_QUALITY_P9 2.0 + #define FXAA_QUALITY_P10 4.0 + #define FXAA_QUALITY_P11 8.0 +#endif + +/*============================================================================ + FXAA QUALITY - EXTREME QUALITY +============================================================================*/ +#if (FXAA_QUALITY_PRESET == 39) + #define FXAA_QUALITY_PS 12 + #define FXAA_QUALITY_P0 1.0 + #define FXAA_QUALITY_P1 1.0 + #define FXAA_QUALITY_P2 1.0 + #define FXAA_QUALITY_P3 1.0 + #define FXAA_QUALITY_P4 1.0 + #define FXAA_QUALITY_P5 1.5 + #define FXAA_QUALITY_P6 2.0 + #define FXAA_QUALITY_P7 2.0 + #define FXAA_QUALITY_P8 2.0 + #define FXAA_QUALITY_P9 2.0 + #define FXAA_QUALITY_P10 4.0 + #define FXAA_QUALITY_P11 8.0 +#endif + + + +/*============================================================================ + + API PORTING + +============================================================================*/ +#if (FXAA_GLSL_120 == 1) || (FXAA_GLSL_130 == 1) + #define FxaaBool bool + #define FxaaDiscard discard + #define FxaaFloat float + #define FxaaFloat2 vec2 + #define FxaaFloat3 vec3 + #define FxaaFloat4 vec4 + #define FxaaHalf float + #define FxaaHalf2 vec2 + #define FxaaHalf3 vec3 + #define FxaaHalf4 vec4 + #define FxaaInt2 ivec2 + #define FxaaSat(x) clamp(x, 0.0, 1.0) + #define FxaaTex sampler2D +#else + #define FxaaBool bool + #define FxaaDiscard clip(-1) + #define FxaaFloat float + #define FxaaFloat2 float2 + #define FxaaFloat3 float3 + #define FxaaFloat4 float4 + #define FxaaHalf half + #define FxaaHalf2 half2 + #define FxaaHalf3 half3 + #define FxaaHalf4 half4 + #define FxaaSat(x) saturate(x) +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_GLSL_120 == 1) + // Requires, + // #version 120 + // And at least, + // #extension GL_EXT_gpu_shader4 : enable + // (or set FXAA_FAST_PIXEL_OFFSET 1 to work like DX9) + #define FxaaTexTop(t, p) texture2DLod(t, p, 0.0) + #if (FXAA_FAST_PIXEL_OFFSET == 1) + #define FxaaTexOff(t, p, o, r) texture2DLodOffset(t, p, 0.0, o) + #else + #define FxaaTexOff(t, p, o, r) texture2DLod(t, p + (o * r), 0.0) + #endif + #if (FXAA_GATHER4_ALPHA == 1) + // use #extension GL_ARB_gpu_shader5 : enable + #define FxaaTexAlpha4(t, p) textureGather(t, p, 3) + #define FxaaTexOffAlpha4(t, p, o) textureGatherOffset(t, p, o, 3) + #define FxaaTexGreen4(t, p) textureGather(t, p, 1) + #define FxaaTexOffGreen4(t, p, o) textureGatherOffset(t, p, o, 1) + #endif +#endif +/*--------------------------------------------------------------------------*/ +#if (FXAA_GLSL_130 == 1) + // Requires "#version 130" or better + #define FxaaTexTop(t, p) textureLod(t, p, 0.0) + #define FxaaTexOff(t, p, o, r) textureLodOffset(t, p, 0.0, o) + #if (FXAA_GATHER4_ALPHA == 1) + // use #extension GL_ARB_gpu_shader5 : enable + #define FxaaTexAlpha4(t, p) textureGather(t, p, 3) + #define FxaaTexOffAlpha4(t, p, o) textureGatherOffset(t, p, o, 3) + #define FxaaTexGreen4(t, p) textureGather(t, p, 1) + #define FxaaTexOffGreen4(t, p, o) textureGatherOffset(t, p, o, 1) + #endif +#endif + + +/*============================================================================ + GREEN AS LUMA OPTION SUPPORT FUNCTION +============================================================================*/ +#if (FXAA_GREEN_AS_LUMA == 0) + FxaaFloat FxaaLuma(FxaaFloat4 rgba) { return rgba.w; } +#else + FxaaFloat FxaaLuma(FxaaFloat4 rgba) { return rgba.y; } +#endif + + + + +/*============================================================================ + + FXAA3 QUALITY - PC + +============================================================================*/ +#if (FXAA_PC == 1) +/*--------------------------------------------------------------------------*/ +FxaaFloat4 FxaaPixelShader( + // + // Use noperspective interpolation here (turn off perspective interpolation). + // {xy} = center of pixel + FxaaFloat2 pos, + // + // Used only for FXAA Console, and not used on the 360 version. + // Use noperspective interpolation here (turn off perspective interpolation). + // {xy_} = upper left of pixel + // {_zw} = lower right of pixel + FxaaFloat4 fxaaConsolePosPos, + // + // Input color texture. + // {rgb_} = color in linear or perceptual color space + // if (FXAA_GREEN_AS_LUMA == 0) + // {__a} = luma in perceptual color space (not linear) + FxaaTex tex, + // + // Only used on the optimized 360 version of FXAA Console. + // For everything but 360, just use the same input here as for "tex". + // For 360, same texture, just alias with a 2nd sampler. + // This sampler needs to have an exponent bias of -1. + FxaaTex fxaaConsole360TexExpBiasNegOne, + // + // Only used on the optimized 360 version of FXAA Console. + // For everything but 360, just use the same input here as for "tex". + // For 360, same texture, just alias with a 3nd sampler. + // This sampler needs to have an exponent bias of -2. + FxaaTex fxaaConsole360TexExpBiasNegTwo, + // + // Only used on FXAA Quality. + // This must be from a constant/uniform. + // {x_} = 1.0/screenWidthInPixels + // {_y} = 1.0/screenHeightInPixels + FxaaFloat2 fxaaQualityRcpFrame, + // + // Only used on FXAA Console. + // This must be from a constant/uniform. + // This effects sub-pixel AA quality and inversely sharpness. + // Where N ranges between, + // N = 0.50 (default) + // N = 0.33 (sharper) + // {x__} = -N/screenWidthInPixels + // {_y_} = -N/screenHeightInPixels + // {_z_} = N/screenWidthInPixels + // {__w} = N/screenHeightInPixels + FxaaFloat4 fxaaConsoleRcpFrameOpt, + // + // Only used on FXAA Console. + // Not used on 360, but used on PS3 and PC. + // This must be from a constant/uniform. + // {x__} = -2.0/screenWidthInPixels + // {_y_} = -2.0/screenHeightInPixels + // {_z_} = 2.0/screenWidthInPixels + // {__w} = 2.0/screenHeightInPixels + FxaaFloat4 fxaaConsoleRcpFrameOpt2, + // + // Only used on FXAA Console. + // Only used on 360 in place of fxaaConsoleRcpFrameOpt2. + // This must be from a constant/uniform. + // {x__} = 8.0/screenWidthInPixels + // {_y_} = 8.0/screenHeightInPixels + // {_z_} = -4.0/screenWidthInPixels + // {__w} = -4.0/screenHeightInPixels + FxaaFloat4 fxaaConsole360RcpFrameOpt2, + // + // Only used on FXAA Quality. + // This used to be the FXAA_QUALITY_SUBPIX define. + // It is here now to allow easier tuning. + // Choose the amount of sub-pixel aliasing removal. + // This can effect sharpness. + // 1.00 - upper limit (softer) + // 0.75 - default amount of filtering + // 0.50 - lower limit (sharper, less sub-pixel aliasing removal) + // 0.25 - almost off + // 0.00 - completely off + FxaaFloat fxaaQualitySubpix, + // + // Only used on FXAA Quality. + // This used to be the FXAA_QUALITY_EDGE_THRESHOLD define. + // It is here now to allow easier tuning. + // The minimum amount of local contrast required to apply algorithm. + // 0.333 - too little (faster) + // 0.250 - low quality + // 0.166 - default + // 0.125 - high quality + // 0.063 - overkill (slower) + FxaaFloat fxaaQualityEdgeThreshold, + // + // Only used on FXAA Quality. + // This used to be the FXAA_QUALITY_EDGE_THRESHOLD_MIN define. + // It is here now to allow easier tuning. + // Trims the algorithm from processing darks. + // 0.0833 - upper limit (default, the start of visible unfiltered edges) + // 0.0625 - high quality (faster) + // 0.0312 - visible limit (slower) + // Special notes when using FXAA_GREEN_AS_LUMA, + // Likely want to set this to zero. + // As colors that are mostly not-green + // will appear very dark in the green channel! + // Tune by looking at mostly non-green content, + // then start at zero and increase until aliasing is a problem. + FxaaFloat fxaaQualityEdgeThresholdMin, + // + // Only used on FXAA Console. + // This used to be the FXAA_CONSOLE_EDGE_SHARPNESS define. + // It is here now to allow easier tuning. + // This does not effect PS3, as this needs to be compiled in. + // Use FXAA_CONSOLE_PS3_EDGE_SHARPNESS for PS3. + // Due to the PS3 being ALU bound, + // there are only three safe values here: 2 and 4 and 8. + // These options use the shaders ability to a free *|/ by 2|4|8. + // For all other platforms can be a non-power of two. + // 8.0 is sharper (default!!!) + // 4.0 is softer + // 2.0 is really soft (good only for vector graphics inputs) + FxaaFloat fxaaConsoleEdgeSharpness, + // + // Only used on FXAA Console. + // This used to be the FXAA_CONSOLE_EDGE_THRESHOLD define. + // It is here now to allow easier tuning. + // This does not effect PS3, as this needs to be compiled in. + // Use FXAA_CONSOLE_PS3_EDGE_THRESHOLD for PS3. + // Due to the PS3 being ALU bound, + // there are only two safe values here: 1/4 and 1/8. + // These options use the shaders ability to a free *|/ by 2|4|8. + // The console setting has a different mapping than the quality setting. + // Other platforms can use other values. + // 0.125 leaves less aliasing, but is softer (default!!!) + // 0.25 leaves more aliasing, and is sharper + FxaaFloat fxaaConsoleEdgeThreshold, + // + // Only used on FXAA Console. + // This used to be the FXAA_CONSOLE_EDGE_THRESHOLD_MIN define. + // It is here now to allow easier tuning. + // Trims the algorithm from processing darks. + // The console setting has a different mapping than the quality setting. + // This only applies when FXAA_EARLY_EXIT is 1. + // This does not apply to PS3, + // PS3 was simplified to avoid more shader instructions. + // 0.06 - faster but more aliasing in darks + // 0.05 - default + // 0.04 - slower and less aliasing in darks + // Special notes when using FXAA_GREEN_AS_LUMA, + // Likely want to set this to zero. + // As colors that are mostly not-green + // will appear very dark in the green channel! + // Tune by looking at mostly non-green content, + // then start at zero and increase until aliasing is a problem. + FxaaFloat fxaaConsoleEdgeThresholdMin, + // + // Extra constants for 360 FXAA Console only. + // Use zeros or anything else for other platforms. + // These must be in physical constant registers and NOT immedates. + // Immedates will result in compiler un-optimizing. + // {xyzw} = float4(1.0, -1.0, 0.25, -0.25) + FxaaFloat4 fxaaConsole360ConstDir +) { +/*--------------------------------------------------------------------------*/ + FxaaFloat2 posM; + posM.x = pos.x; + posM.y = pos.y; + #if (FXAA_GATHER4_ALPHA == 1) + #if (FXAA_DISCARD == 0) + FxaaFloat4 rgbyM = FxaaTexTop(tex, posM); + #if (FXAA_GREEN_AS_LUMA == 0) + #define lumaM rgbyM.w + #else + #define lumaM rgbyM.y + #endif + #endif + #if (FXAA_GREEN_AS_LUMA == 0) + FxaaFloat4 luma4A = FxaaTexAlpha4(tex, posM); + FxaaFloat4 luma4B = FxaaTexOffAlpha4(tex, posM, FxaaInt2(-1, -1)); + #else + FxaaFloat4 luma4A = FxaaTexGreen4(tex, posM); + FxaaFloat4 luma4B = FxaaTexOffGreen4(tex, posM, FxaaInt2(-1, -1)); + #endif + #if (FXAA_DISCARD == 1) + #define lumaM luma4A.w + #endif + #define lumaE luma4A.z + #define lumaS luma4A.x + #define lumaSE luma4A.y + #define lumaNW luma4B.w + #define lumaN luma4B.z + #define lumaW luma4B.x + #else + FxaaFloat4 rgbyM = FxaaTexTop(tex, posM); + #if (FXAA_GREEN_AS_LUMA == 0) + #define lumaM rgbyM.w + #else + #define lumaM rgbyM.y + #endif + FxaaFloat lumaS = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2( 0, 1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaE = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2( 1, 0), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaN = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2( 0,-1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaW = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2(-1, 0), fxaaQualityRcpFrame.xy)); + #endif +/*--------------------------------------------------------------------------*/ + FxaaFloat maxSM = max(lumaS, lumaM); + FxaaFloat minSM = min(lumaS, lumaM); + FxaaFloat maxESM = max(lumaE, maxSM); + FxaaFloat minESM = min(lumaE, minSM); + FxaaFloat maxWN = max(lumaN, lumaW); + FxaaFloat minWN = min(lumaN, lumaW); + FxaaFloat rangeMax = max(maxWN, maxESM); + FxaaFloat rangeMin = min(minWN, minESM); + FxaaFloat rangeMaxScaled = rangeMax * fxaaQualityEdgeThreshold; + FxaaFloat range = rangeMax - rangeMin; + FxaaFloat rangeMaxClamped = max(fxaaQualityEdgeThresholdMin, rangeMaxScaled); + FxaaBool earlyExit = range < rangeMaxClamped; +/*--------------------------------------------------------------------------*/ + if(earlyExit) + #if (FXAA_DISCARD == 1) + FxaaDiscard; + #else + return rgbyM; + #endif +/*--------------------------------------------------------------------------*/ + #if (FXAA_GATHER4_ALPHA == 0) + FxaaFloat lumaNW = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2(-1,-1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaSE = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2( 1, 1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaNE = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2( 1,-1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaSW = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2(-1, 1), fxaaQualityRcpFrame.xy)); + #else + FxaaFloat lumaNE = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2(1, -1), fxaaQualityRcpFrame.xy)); + FxaaFloat lumaSW = FxaaLuma(FxaaTexOff(tex, posM, FxaaInt2(-1, 1), fxaaQualityRcpFrame.xy)); + #endif +/*--------------------------------------------------------------------------*/ + FxaaFloat lumaNS = lumaN + lumaS; + FxaaFloat lumaWE = lumaW + lumaE; + FxaaFloat subpixRcpRange = 1.0/range; + FxaaFloat subpixNSWE = lumaNS + lumaWE; + FxaaFloat edgeHorz1 = (-2.0 * lumaM) + lumaNS; + FxaaFloat edgeVert1 = (-2.0 * lumaM) + lumaWE; +/*--------------------------------------------------------------------------*/ + FxaaFloat lumaNESE = lumaNE + lumaSE; + FxaaFloat lumaNWNE = lumaNW + lumaNE; + FxaaFloat edgeHorz2 = (-2.0 * lumaE) + lumaNESE; + FxaaFloat edgeVert2 = (-2.0 * lumaN) + lumaNWNE; +/*--------------------------------------------------------------------------*/ + FxaaFloat lumaNWSW = lumaNW + lumaSW; + FxaaFloat lumaSWSE = lumaSW + lumaSE; + FxaaFloat edgeHorz4 = (abs(edgeHorz1) * 2.0) + abs(edgeHorz2); + FxaaFloat edgeVert4 = (abs(edgeVert1) * 2.0) + abs(edgeVert2); + FxaaFloat edgeHorz3 = (-2.0 * lumaW) + lumaNWSW; + FxaaFloat edgeVert3 = (-2.0 * lumaS) + lumaSWSE; + FxaaFloat edgeHorz = abs(edgeHorz3) + edgeHorz4; + FxaaFloat edgeVert = abs(edgeVert3) + edgeVert4; +/*--------------------------------------------------------------------------*/ + FxaaFloat subpixNWSWNESE = lumaNWSW + lumaNESE; + FxaaFloat lengthSign = fxaaQualityRcpFrame.x; + FxaaBool horzSpan = edgeHorz >= edgeVert; + FxaaFloat subpixA = subpixNSWE * 2.0 + subpixNWSWNESE; +/*--------------------------------------------------------------------------*/ + if(!horzSpan) lumaN = lumaW; + if(!horzSpan) lumaS = lumaE; + if(horzSpan) lengthSign = fxaaQualityRcpFrame.y; + FxaaFloat subpixB = (subpixA * (1.0/12.0)) - lumaM; +/*--------------------------------------------------------------------------*/ + FxaaFloat gradientN = lumaN - lumaM; + FxaaFloat gradientS = lumaS - lumaM; + FxaaFloat lumaNN = lumaN + lumaM; + FxaaFloat lumaSS = lumaS + lumaM; + FxaaBool pairN = abs(gradientN) >= abs(gradientS); + FxaaFloat gradient = max(abs(gradientN), abs(gradientS)); + if(pairN) lengthSign = -lengthSign; + FxaaFloat subpixC = FxaaSat(abs(subpixB) * subpixRcpRange); +/*--------------------------------------------------------------------------*/ + FxaaFloat2 posB; + posB.x = posM.x; + posB.y = posM.y; + FxaaFloat2 offNP; + offNP.x = (!horzSpan) ? 0.0 : fxaaQualityRcpFrame.x; + offNP.y = ( horzSpan) ? 0.0 : fxaaQualityRcpFrame.y; + if(!horzSpan) posB.x += lengthSign * 0.5; + if( horzSpan) posB.y += lengthSign * 0.5; +/*--------------------------------------------------------------------------*/ + FxaaFloat2 posN; + posN.x = posB.x - offNP.x * FXAA_QUALITY_P0; + posN.y = posB.y - offNP.y * FXAA_QUALITY_P0; + FxaaFloat2 posP; + posP.x = posB.x + offNP.x * FXAA_QUALITY_P0; + posP.y = posB.y + offNP.y * FXAA_QUALITY_P0; + FxaaFloat subpixD = ((-2.0)*subpixC) + 3.0; + FxaaFloat lumaEndN = FxaaLuma(FxaaTexTop(tex, posN)); + FxaaFloat subpixE = subpixC * subpixC; + FxaaFloat lumaEndP = FxaaLuma(FxaaTexTop(tex, posP)); +/*--------------------------------------------------------------------------*/ + if(!pairN) lumaNN = lumaSS; + FxaaFloat gradientScaled = gradient * 1.0/4.0; + FxaaFloat lumaMM = lumaM - lumaNN * 0.5; + FxaaFloat subpixF = subpixD * subpixE; + FxaaBool lumaMLTZero = lumaMM < 0.0; +/*--------------------------------------------------------------------------*/ + lumaEndN -= lumaNN * 0.5; + lumaEndP -= lumaNN * 0.5; + FxaaBool doneN = abs(lumaEndN) >= gradientScaled; + FxaaBool doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P1; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P1; + FxaaBool doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P1; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P1; +/*--------------------------------------------------------------------------*/ + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P2; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P2; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P2; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P2; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY_PS > 3) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P3; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P3; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P3; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P3; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY_PS > 4) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P4; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P4; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P4; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P4; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY_PS > 5) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P5; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P5; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P5; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P5; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY_PS > 6) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P6; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P6; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P6; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P6; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY_PS > 7) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P7; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P7; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P7; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P7; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY_PS > 8) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P8; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P8; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P8; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P8; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY_PS > 9) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P9; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P9; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P9; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P9; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY_PS > 10) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P10; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P10; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P10; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P10; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY_PS > 11) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P11; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P11; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P11; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P11; +/*--------------------------------------------------------------------------*/ + #if (FXAA_QUALITY_PS > 12) + if(doneNP) { + if(!doneN) lumaEndN = FxaaLuma(FxaaTexTop(tex, posN.xy)); + if(!doneP) lumaEndP = FxaaLuma(FxaaTexTop(tex, posP.xy)); + if(!doneN) lumaEndN = lumaEndN - lumaNN * 0.5; + if(!doneP) lumaEndP = lumaEndP - lumaNN * 0.5; + doneN = abs(lumaEndN) >= gradientScaled; + doneP = abs(lumaEndP) >= gradientScaled; + if(!doneN) posN.x -= offNP.x * FXAA_QUALITY_P12; + if(!doneN) posN.y -= offNP.y * FXAA_QUALITY_P12; + doneNP = (!doneN) || (!doneP); + if(!doneP) posP.x += offNP.x * FXAA_QUALITY_P12; + if(!doneP) posP.y += offNP.y * FXAA_QUALITY_P12; +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } + #endif +/*--------------------------------------------------------------------------*/ + } +/*--------------------------------------------------------------------------*/ + FxaaFloat dstN = posM.x - posN.x; + FxaaFloat dstP = posP.x - posM.x; + if(!horzSpan) dstN = posM.y - posN.y; + if(!horzSpan) dstP = posP.y - posM.y; +/*--------------------------------------------------------------------------*/ + FxaaBool goodSpanN = (lumaEndN < 0.0) != lumaMLTZero; + FxaaFloat spanLength = (dstP + dstN); + FxaaBool goodSpanP = (lumaEndP < 0.0) != lumaMLTZero; + FxaaFloat spanLengthRcp = 1.0/spanLength; +/*--------------------------------------------------------------------------*/ + FxaaBool directionN = dstN < dstP; + FxaaFloat dst = min(dstN, dstP); + FxaaBool goodSpan = directionN ? goodSpanN : goodSpanP; + FxaaFloat subpixG = subpixF * subpixF; + FxaaFloat pixelOffset = (dst * (-spanLengthRcp)) + 0.5; + FxaaFloat subpixH = subpixG * fxaaQualitySubpix; +/*--------------------------------------------------------------------------*/ + FxaaFloat pixelOffsetGood = goodSpan ? pixelOffset : 0.0; + FxaaFloat pixelOffsetSubpix = max(pixelOffsetGood, subpixH); + if(!horzSpan) posM.x += pixelOffsetSubpix * lengthSign; + if( horzSpan) posM.y += pixelOffsetSubpix * lengthSign; + #if (FXAA_DISCARD == 1) + return FxaaTexTop(tex, posM); + #else + return FxaaFloat4(FxaaTexTop(tex, posM).xyz, lumaM); + #endif +} +/*==========================================================================*/ +#endif + +vec4 mainImage(vec2 fragCoord) +{ + vec2 rcpFrame = 1./invResolution_data.xy; + vec2 uv2 = fragCoord.xy / invResolution_data.xy; + + float fxaaQualitySubpix = 0.75; // [0..1], default 0.75 + float fxaaQualityEdgeThreshold = 0.166; // [0.125..0.33], default 0.166 + float fxaaQualityEdgeThresholdMin = 0.02;//0.0625; // ? + vec4 dummy4 = vec4(0.0,0.0,0.0,0.0); + float dummy1 = 0.0; + + vec4 col = FxaaPixelShader(uv2, dummy4, + inputImage, inputImage, inputImage, + rcpFrame, dummy4, dummy4, dummy4, + fxaaQualitySubpix, fxaaQualityEdgeThreshold, + fxaaQualityEdgeThresholdMin, + dummy1, dummy1, dummy1, dummy4); + + vec4 fragColor = vec4( col.xyz, 1. ); + + return fragColor; +} + +void main() +{ + ivec2 loc = ivec2(gl_GlobalInvocationID.x * 4, gl_GlobalInvocationID.y * 4); + for(int i = 0; i < 4; i++) + { + for(int j = 0; j < 4; j++) + { + ivec2 texelCoord = ivec2(loc.x + i, loc.y + j); + vec4 outColor = mainImage(texelCoord + vec2(0.5)); + imageStore(imgOutput, texelCoord, outColor); + } + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/Fxaa.spv b/Ryujinx.Graphics.Vulkan/Effects/Shaders/Fxaa.spv new file mode 100644 index 00000000..b466bcb6 Binary files /dev/null and b/Ryujinx.Graphics.Vulkan/Effects/Shaders/Fxaa.spv differ diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaBlend.glsl b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaBlend.glsl new file mode 100644 index 00000000..a518cf25 --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaBlend.glsl @@ -0,0 +1,1404 @@ +#version 430 core +#define SMAA_GLSL_4 1 + +layout (constant_id = 0) const int SMAA_PRESET_LOW = 0; +layout (constant_id = 1) const int SMAA_PRESET_MEDIUM = 0; +layout (constant_id = 2) const int SMAA_PRESET_HIGH = 0; +layout (constant_id = 3) const int SMAA_PRESET_ULTRA = 0; +layout (constant_id = 4) const float METRIC_WIDTH = 1920.0; +layout (constant_id = 5) const float METRIC_HEIGHT = 1080.0; + +#define SMAA_RT_METRICS float4(1.0 / METRIC_WIDTH, 1.0 / METRIC_HEIGHT, METRIC_WIDTH, METRIC_HEIGHT) + +layout (local_size_x = 16, local_size_y = 16) in; +/** + * Copyright (C) 2013 Jorge Jimenez (jorge@iryoku.com) + * Copyright (C) 2013 Jose I. Echevarria (joseignacioechevarria@gmail.com) + * Copyright (C) 2013 Belen Masia (bmasia@unizar.es) + * Copyright (C) 2013 Fernando Navarro (fernandn@microsoft.com) + * Copyright (C) 2013 Diego Gutierrez (diegog@unizar.es) + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * this software and associated documentation files (the "Software"), to deal in + * the Software without restriction, including without limitation the rights to + * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies + * of the Software, and to permit persons to whom the Software is furnished to + * do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. As clarification, there + * is no requirement that the copyright notice and permission be included in + * binary distributions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + + +/** + * _______ ___ ___ ___ ___ + * / || \/ | / \ / \ + * | (---- | \ / | / ^ \ / ^ \ + * \ \ | |\/| | / /_\ \ / /_\ \ + * ----) | | | | | / _____ \ / _____ \ + * |_______/ |__| |__| /__/ \__\ /__/ \__\ + * + * E N H A N C E D + * S U B P I X E L M O R P H O L O G I C A L A N T I A L I A S I N G + * + * http://www.iryoku.com/smaa/ + * + * Hi, welcome aboard! + * + * Here you'll find instructions to get the shader up and running as fast as + * possible. + * + * IMPORTANTE NOTICE: when updating, remember to update both this file and the + * precomputed textures! They may change from version to version. + * + * The shader has three passes, chained together as follows: + * + * |input|------------------ + * v | + * [ SMAA*EdgeDetection ] | + * v | + * |edgesTex| | + * v | + * [ SMAABlendingWeightCalculation ] | + * v | + * |blendTex| | + * v | + * [ SMAANeighborhoodBlending ] <------ + * v + * |output| + * + * Note that each [pass] has its own vertex and pixel shader. Remember to use + * oversized triangles instead of quads to avoid overshading along the + * diagonal. + * + * You've three edge detection methods to choose from: luma, color or depth. + * They represent different quality/performance and anti-aliasing/sharpness + * tradeoffs, so our recommendation is for you to choose the one that best + * suits your particular scenario: + * + * - Depth edge detection is usually the fastest but it may miss some edges. + * + * - Luma edge detection is usually more expensive than depth edge detection, + * but catches visible edges that depth edge detection can miss. + * + * - Color edge detection is usually the most expensive one but catches + * chroma-only edges. + * + * For quickstarters: just use luma edge detection. + * + * The general advice is to not rush the integration process and ensure each + * step is done correctly (don't try to integrate SMAA T2x with predicated edge + * detection from the start!). Ok then, let's go! + * + * 1. The first step is to create two RGBA temporal render targets for holding + * |edgesTex| and |blendTex|. + * + * In DX10 or DX11, you can use a RG render target for the edges texture. + * In the case of NVIDIA GPUs, using RG render targets seems to actually be + * slower. + * + * On the Xbox 360, you can use the same render target for resolving both + * |edgesTex| and |blendTex|, as they aren't needed simultaneously. + * + * 2. Both temporal render targets |edgesTex| and |blendTex| must be cleared + * each frame. Do not forget to clear the alpha channel! + * + * 3. The next step is loading the two supporting precalculated textures, + * 'areaTex' and 'searchTex'. You'll find them in the 'Textures' folder as + * C++ headers, and also as regular DDS files. They'll be needed for the + * 'SMAABlendingWeightCalculation' pass. + * + * If you use the C++ headers, be sure to load them in the format specified + * inside of them. + * + * You can also compress 'areaTex' and 'searchTex' using BC5 and BC4 + * respectively, if you have that option in your content processor pipeline. + * When compressing then, you get a non-perceptible quality decrease, and a + * marginal performance increase. + * + * 4. All samplers must be set to linear filtering and clamp. + * + * After you get the technique working, remember that 64-bit inputs have + * half-rate linear filtering on GCN. + * + * If SMAA is applied to 64-bit color buffers, switching to point filtering + * when accesing them will increase the performance. Search for + * 'SMAASamplePoint' to see which textures may benefit from point + * filtering, and where (which is basically the color input in the edge + * detection and resolve passes). + * + * 5. All texture reads and buffer writes must be non-sRGB, with the exception + * of the input read and the output write in + * 'SMAANeighborhoodBlending' (and only in this pass!). If sRGB reads in + * this last pass are not possible, the technique will work anyway, but + * will perform antialiasing in gamma space. + * + * IMPORTANT: for best results the input read for the color/luma edge + * detection should *NOT* be sRGB. + * + * 6. Before including SMAA.h you'll have to setup the render target metrics, + * the target and any optional configuration defines. Optionally you can + * use a preset. + * + * You have the following targets available: + * SMAA_HLSL_3 + * SMAA_HLSL_4 + * SMAA_HLSL_4_1 + * SMAA_GLSL_3 * + * SMAA_GLSL_4 * + * + * * (See SMAA_INCLUDE_VS and SMAA_INCLUDE_PS below). + * + * And four presets: + * SMAA_PRESET_LOW (%60 of the quality) + * SMAA_PRESET_MEDIUM (%80 of the quality) + * SMAA_PRESET_HIGH (%95 of the quality) + * SMAA_PRESET_ULTRA (%99 of the quality) + * + * For example: + * #define SMAA_RT_METRICS float4(1.0 / 1280.0, 1.0 / 720.0, 1280.0, 720.0) + * #define SMAA_HLSL_4 + * #define SMAA_PRESET_HIGH + * #include "SMAA.h" + * + * Note that SMAA_RT_METRICS doesn't need to be a macro, it can be a + * uniform variable. The code is designed to minimize the impact of not + * using a constant value, but it is still better to hardcode it. + * + * Depending on how you encoded 'areaTex' and 'searchTex', you may have to + * add (and customize) the following defines before including SMAA.h: + * #define SMAA_AREATEX_SELECT(sample) sample.rg + * #define SMAA_SEARCHTEX_SELECT(sample) sample.r + * + * If your engine is already using porting macros, you can define + * SMAA_CUSTOM_SL, and define the porting functions by yourself. + * + * 7. Then, you'll have to setup the passes as indicated in the scheme above. + * You can take a look into SMAA.fx, to see how we did it for our demo. + * Checkout the function wrappers, you may want to copy-paste them! + * + * 8. It's recommended to validate the produced |edgesTex| and |blendTex|. + * You can use a screenshot from your engine to compare the |edgesTex| + * and |blendTex| produced inside of the engine with the results obtained + * with the reference demo. + * + * 9. After you get the last pass to work, it's time to optimize. You'll have + * to initialize a stencil buffer in the first pass (discard is already in + * the code), then mask execution by using it the second pass. The last + * pass should be executed in all pixels. + * + * + * After this point you can choose to enable predicated thresholding, + * temporal supersampling and motion blur integration: + * + * a) If you want to use predicated thresholding, take a look into + * SMAA_PREDICATION; you'll need to pass an extra texture in the edge + * detection pass. + * + * b) If you want to enable temporal supersampling (SMAA T2x): + * + * 1. The first step is to render using subpixel jitters. I won't go into + * detail, but it's as simple as moving each vertex position in the + * vertex shader, you can check how we do it in our DX10 demo. + * + * 2. Then, you must setup the temporal resolve. You may want to take a look + * into SMAAResolve for resolving 2x modes. After you get it working, you'll + * probably see ghosting everywhere. But fear not, you can enable the + * CryENGINE temporal reprojection by setting the SMAA_REPROJECTION macro. + * Check out SMAA_DECODE_VELOCITY if your velocity buffer is encoded. + * + * 3. The next step is to apply SMAA to each subpixel jittered frame, just as + * done for 1x. + * + * 4. At this point you should already have something usable, but for best + * results the proper area textures must be set depending on current jitter. + * For this, the parameter 'subsampleIndices' of + * 'SMAABlendingWeightCalculationPS' must be set as follows, for our T2x + * mode: + * + * @SUBSAMPLE_INDICES + * + * | S# | Camera Jitter | subsampleIndices | + * +----+------------------+---------------------+ + * | 0 | ( 0.25, -0.25) | float4(1, 1, 1, 0) | + * | 1 | (-0.25, 0.25) | float4(2, 2, 2, 0) | + * + * These jitter positions assume a bottom-to-top y axis. S# stands for the + * sample number. + * + * More information about temporal supersampling here: + * http://iryoku.com/aacourse/downloads/13-Anti-Aliasing-Methods-in-CryENGINE-3.pdf + * + * c) If you want to enable spatial multisampling (SMAA S2x): + * + * 1. The scene must be rendered using MSAA 2x. The MSAA 2x buffer must be + * created with: + * - DX10: see below (*) + * - DX10.1: D3D10_STANDARD_MULTISAMPLE_PATTERN or + * - DX11: D3D11_STANDARD_MULTISAMPLE_PATTERN + * + * This allows to ensure that the subsample order matches the table in + * @SUBSAMPLE_INDICES. + * + * (*) In the case of DX10, we refer the reader to: + * - SMAA::detectMSAAOrder and + * - SMAA::msaaReorder + * + * These functions allow to match the standard multisample patterns by + * detecting the subsample order for a specific GPU, and reordering + * them appropriately. + * + * 2. A shader must be run to output each subsample into a separate buffer + * (DX10 is required). You can use SMAASeparate for this purpose, or just do + * it in an existing pass (for example, in the tone mapping pass, which has + * the advantage of feeding tone mapped subsamples to SMAA, which will yield + * better results). + * + * 3. The full SMAA 1x pipeline must be run for each separated buffer, storing + * the results in the final buffer. The second run should alpha blend with + * the existing final buffer using a blending factor of 0.5. + * 'subsampleIndices' must be adjusted as in the SMAA T2x case (see point + * b). + * + * d) If you want to enable temporal supersampling on top of SMAA S2x + * (which actually is SMAA 4x): + * + * 1. SMAA 4x consists on temporally jittering SMAA S2x, so the first step is + * to calculate SMAA S2x for current frame. In this case, 'subsampleIndices' + * must be set as follows: + * + * | F# | S# | Camera Jitter | Net Jitter | subsampleIndices | + * +----+----+--------------------+-------------------+----------------------+ + * | 0 | 0 | ( 0.125, 0.125) | ( 0.375, -0.125) | float4(5, 3, 1, 3) | + * | 0 | 1 | ( 0.125, 0.125) | (-0.125, 0.375) | float4(4, 6, 2, 3) | + * +----+----+--------------------+-------------------+----------------------+ + * | 1 | 2 | (-0.125, -0.125) | ( 0.125, -0.375) | float4(3, 5, 1, 4) | + * | 1 | 3 | (-0.125, -0.125) | (-0.375, 0.125) | float4(6, 4, 2, 4) | + * + * These jitter positions assume a bottom-to-top y axis. F# stands for the + * frame number. S# stands for the sample number. + * + * 2. After calculating SMAA S2x for current frame (with the new subsample + * indices), previous frame must be reprojected as in SMAA T2x mode (see + * point b). + * + * e) If motion blur is used, you may want to do the edge detection pass + * together with motion blur. This has two advantages: + * + * 1. Pixels under heavy motion can be omitted from the edge detection process. + * For these pixels we can just store "no edge", as motion blur will take + * care of them. + * 2. The center pixel tap is reused. + * + * Note that in this case depth testing should be used instead of stenciling, + * as we have to write all the pixels in the motion blur pass. + * + * That's it! + */ + +//----------------------------------------------------------------------------- +// SMAA Presets + +/** + * Note that if you use one of these presets, the following configuration + * macros will be ignored if set in the "Configurable Defines" section. + */ + +#if defined(SMAA_PRESET_LOW) +#define SMAA_THRESHOLD 0.15 +#define SMAA_MAX_SEARCH_STEPS 4 +#define SMAA_DISABLE_DIAG_DETECTION +#define SMAA_DISABLE_CORNER_DETECTION +#elif defined(SMAA_PRESET_MEDIUM) +#define SMAA_THRESHOLD 0.1 +#define SMAA_MAX_SEARCH_STEPS 8 +#define SMAA_DISABLE_DIAG_DETECTION +#define SMAA_DISABLE_CORNER_DETECTION +#elif defined(SMAA_PRESET_HIGH) +#define SMAA_THRESHOLD 0.1 +#define SMAA_MAX_SEARCH_STEPS 16 +#define SMAA_MAX_SEARCH_STEPS_DIAG 8 +#define SMAA_CORNER_ROUNDING 25 +#elif defined(SMAA_PRESET_ULTRA) +#define SMAA_THRESHOLD 0.05 +#define SMAA_MAX_SEARCH_STEPS 32 +#define SMAA_MAX_SEARCH_STEPS_DIAG 16 +#define SMAA_CORNER_ROUNDING 25 +#endif + +//----------------------------------------------------------------------------- +// Configurable Defines + +/** + * SMAA_THRESHOLD specifies the threshold or sensitivity to edges. + * Lowering this value you will be able to detect more edges at the expense of + * performance. + * + * Range: [0, 0.5] + * 0.1 is a reasonable value, and allows to catch most visible edges. + * 0.05 is a rather overkill value, that allows to catch 'em all. + * + * If temporal supersampling is used, 0.2 could be a reasonable value, as low + * contrast edges are properly filtered by just 2x. + */ +#ifndef SMAA_THRESHOLD +#define SMAA_THRESHOLD 0.1 +#endif + +/** + * SMAA_DEPTH_THRESHOLD specifies the threshold for depth edge detection. + * + * Range: depends on the depth range of the scene. + */ +#ifndef SMAA_DEPTH_THRESHOLD +#define SMAA_DEPTH_THRESHOLD (0.1 * SMAA_THRESHOLD) +#endif + +/** + * SMAA_MAX_SEARCH_STEPS specifies the maximum steps performed in the + * horizontal/vertical pattern searches, at each side of the pixel. + * + * In number of pixels, it's actually the double. So the maximum line length + * perfectly handled by, for example 16, is 64 (by perfectly, we meant that + * longer lines won't look as good, but still antialiased). + * + * Range: [0, 112] + */ +#ifndef SMAA_MAX_SEARCH_STEPS +#define SMAA_MAX_SEARCH_STEPS 16 +#endif + +/** + * SMAA_MAX_SEARCH_STEPS_DIAG specifies the maximum steps performed in the + * diagonal pattern searches, at each side of the pixel. In this case we jump + * one pixel at time, instead of two. + * + * Range: [0, 20] + * + * On high-end machines it is cheap (between a 0.8x and 0.9x slower for 16 + * steps), but it can have a significant impact on older machines. + * + * Define SMAA_DISABLE_DIAG_DETECTION to disable diagonal processing. + */ +#ifndef SMAA_MAX_SEARCH_STEPS_DIAG +#define SMAA_MAX_SEARCH_STEPS_DIAG 8 +#endif + +/** + * SMAA_CORNER_ROUNDING specifies how much sharp corners will be rounded. + * + * Range: [0, 100] + * + * Define SMAA_DISABLE_CORNER_DETECTION to disable corner processing. + */ +#ifndef SMAA_CORNER_ROUNDING +#define SMAA_CORNER_ROUNDING 25 +#endif + +/** + * If there is an neighbor edge that has SMAA_LOCAL_CONTRAST_FACTOR times + * bigger contrast than current edge, current edge will be discarded. + * + * This allows to eliminate spurious crossing edges, and is based on the fact + * that, if there is too much contrast in a direction, that will hide + * perceptually contrast in the other neighbors. + */ +#ifndef SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR +#define SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR 2.0 +#endif + +/** + * Predicated thresholding allows to better preserve texture details and to + * improve performance, by decreasing the number of detected edges using an + * additional buffer like the light accumulation buffer, object ids or even the + * depth buffer (the depth buffer usage may be limited to indoor or short range + * scenes). + * + * It locally decreases the luma or color threshold if an edge is found in an + * additional buffer (so the global threshold can be higher). + * + * This method was developed by Playstation EDGE MLAA team, and used in + * Killzone 3, by using the light accumulation buffer. More information here: + * http://iryoku.com/aacourse/downloads/06-MLAA-on-PS3.pptx + */ +#ifndef SMAA_PREDICATION +#define SMAA_PREDICATION 0 +#endif + +/** + * Threshold to be used in the additional predication buffer. + * + * Range: depends on the input, so you'll have to find the magic number that + * works for you. + */ +#ifndef SMAA_PREDICATION_THRESHOLD +#define SMAA_PREDICATION_THRESHOLD 0.01 +#endif + +/** + * How much to scale the global threshold used for luma or color edge + * detection when using predication. + * + * Range: [1, 5] + */ +#ifndef SMAA_PREDICATION_SCALE +#define SMAA_PREDICATION_SCALE 2.0 +#endif + +/** + * How much to locally decrease the threshold. + * + * Range: [0, 1] + */ +#ifndef SMAA_PREDICATION_STRENGTH +#define SMAA_PREDICATION_STRENGTH 0.4 +#endif + +/** + * Temporal reprojection allows to remove ghosting artifacts when using + * temporal supersampling. We use the CryEngine 3 method which also introduces + * velocity weighting. This feature is of extreme importance for totally + * removing ghosting. More information here: + * http://iryoku.com/aacourse/downloads/13-Anti-Aliasing-Methods-in-CryENGINE-3.pdf + * + * Note that you'll need to setup a velocity buffer for enabling reprojection. + * For static geometry, saving the previous depth buffer is a viable + * alternative. + */ +#ifndef SMAA_REPROJECTION +#define SMAA_REPROJECTION 0 +#endif + +/** + * SMAA_REPROJECTION_WEIGHT_SCALE controls the velocity weighting. It allows to + * remove ghosting trails behind the moving object, which are not removed by + * just using reprojection. Using low values will exhibit ghosting, while using + * high values will disable temporal supersampling under motion. + * + * Behind the scenes, velocity weighting removes temporal supersampling when + * the velocity of the subsamples differs (meaning they are different objects). + * + * Range: [0, 80] + */ +#ifndef SMAA_REPROJECTION_WEIGHT_SCALE +#define SMAA_REPROJECTION_WEIGHT_SCALE 30.0 +#endif + +/** + * On some compilers, discard cannot be used in vertex shaders. Thus, they need + * to be compiled separately. + */ +#ifndef SMAA_INCLUDE_VS +#define SMAA_INCLUDE_VS 1 +#endif +#ifndef SMAA_INCLUDE_PS +#define SMAA_INCLUDE_PS 1 +#endif + +//----------------------------------------------------------------------------- +// Texture Access Defines + +#ifndef SMAA_AREATEX_SELECT +#if defined(SMAA_HLSL_3) +#define SMAA_AREATEX_SELECT(sample) sample.ra +#else +#define SMAA_AREATEX_SELECT(sample) sample.rg +#endif +#endif + +#ifndef SMAA_SEARCHTEX_SELECT +#define SMAA_SEARCHTEX_SELECT(sample) sample.r +#endif + +#ifndef SMAA_DECODE_VELOCITY +#define SMAA_DECODE_VELOCITY(sample) sample.rg +#endif + +//----------------------------------------------------------------------------- +// Non-Configurable Defines + +#define SMAA_AREATEX_MAX_DISTANCE 16 +#define SMAA_AREATEX_MAX_DISTANCE_DIAG 20 +#define SMAA_AREATEX_PIXEL_SIZE (1.0 / float2(160.0, 560.0)) +#define SMAA_AREATEX_SUBTEX_SIZE (1.0 / 7.0) +#define SMAA_SEARCHTEX_SIZE float2(66.0, 33.0) +#define SMAA_SEARCHTEX_PACKED_SIZE float2(64.0, 16.0) +#define SMAA_CORNER_ROUNDING_NORM (float(SMAA_CORNER_ROUNDING) / 100.0) + +//----------------------------------------------------------------------------- +// Porting Functions + +#if defined(SMAA_HLSL_3) +#define SMAATexture2D(tex) sampler2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) tex2Dlod(tex, float4(coord, 0.0, 0.0)) +#define SMAASampleLevelZeroPoint(tex, coord) tex2Dlod(tex, float4(coord, 0.0, 0.0)) +#define SMAASampleLevelZeroOffset(tex, coord, offset) tex2Dlod(tex, float4(coord + offset * SMAA_RT_METRICS.xy, 0.0, 0.0)) +#define SMAASample(tex, coord) tex2D(tex, coord) +#define SMAASamplePoint(tex, coord) tex2D(tex, coord) +#define SMAASampleOffset(tex, coord, offset) tex2D(tex, coord + offset * SMAA_RT_METRICS.xy) +#define SMAA_FLATTEN [flatten] +#define SMAA_BRANCH [branch] +#endif +#if defined(SMAA_HLSL_4) || defined(SMAA_HLSL_4_1) +SamplerState LinearSampler { Filter = MIN_MAG_LINEAR_MIP_POINT; AddressU = Clamp; AddressV = Clamp; }; +SamplerState PointSampler { Filter = MIN_MAG_MIP_POINT; AddressU = Clamp; AddressV = Clamp; }; +#define SMAATexture2D(tex) Texture2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) tex.SampleLevel(LinearSampler, coord, 0) +#define SMAASampleLevelZeroPoint(tex, coord) tex.SampleLevel(PointSampler, coord, 0) +#define SMAASampleLevelZeroOffset(tex, coord, offset) tex.SampleLevel(LinearSampler, coord, 0, offset) +#define SMAASample(tex, coord) tex.Sample(LinearSampler, coord) +#define SMAASamplePoint(tex, coord) tex.Sample(PointSampler, coord) +#define SMAASampleOffset(tex, coord, offset) tex.Sample(LinearSampler, coord, offset) +#define SMAA_FLATTEN [flatten] +#define SMAA_BRANCH [branch] +#define SMAATexture2DMS2(tex) Texture2DMS tex +#define SMAALoad(tex, pos, sample) tex.Load(pos, sample) +#if defined(SMAA_HLSL_4_1) +#define SMAAGather(tex, coord) tex.Gather(LinearSampler, coord, 0) +#endif +#endif +#if defined(SMAA_GLSL_3) || defined(SMAA_GLSL_4) +#define SMAATexture2D(tex) sampler2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) textureLod(tex, coord, 0.0) +#define SMAASampleLevelZeroPoint(tex, coord) textureLod(tex, coord, 0.0) +#define SMAASampleLevelZeroOffset(tex, coord, offset) textureLodOffset(tex, coord, 0.0, offset) +#define SMAASample(tex, coord) texture(tex, coord) +#define SMAASamplePoint(tex, coord) texture(tex, coord) +#define SMAASampleOffset(tex, coord, offset) texture(tex, coord, offset) +#define SMAA_FLATTEN +#define SMAA_BRANCH +#define lerp(a, b, t) mix(a, b, t) +#define saturate(a) clamp(a, 0.0, 1.0) +#if defined(SMAA_GLSL_4) +#define mad(a, b, c) fma(a, b, c) +#define SMAAGather(tex, coord) textureGather(tex, coord) +#else +#define mad(a, b, c) (a * b + c) +#endif +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define int2 ivec2 +#define int3 ivec3 +#define int4 ivec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#endif + +#if !defined(SMAA_HLSL_3) && !defined(SMAA_HLSL_4) && !defined(SMAA_HLSL_4_1) && !defined(SMAA_GLSL_3) && !defined(SMAA_GLSL_4) && !defined(SMAA_CUSTOM_SL) +#error you must define the shading language: SMAA_HLSL_*, SMAA_GLSL_* or SMAA_CUSTOM_SL +#endif + +//----------------------------------------------------------------------------- +// Misc functions + +/** + * Gathers current pixel, and the top-left neighbors. + */ +float3 SMAAGatherNeighbours(float2 texcoord, + float4 offset[3], + SMAATexture2D(tex)) { + #ifdef SMAAGather + return SMAAGather(tex, texcoord + SMAA_RT_METRICS.xy * float2(-0.5, -0.5)).grb; + #else + float P = SMAASamplePoint(tex, texcoord).r; + float Pleft = SMAASamplePoint(tex, offset[0].xy).r; + float Ptop = SMAASamplePoint(tex, offset[0].zw).r; + return float3(P, Pleft, Ptop); + #endif +} + +/** + * Adjusts the threshold by means of predication. + */ +float2 SMAACalculatePredicatedThreshold(float2 texcoord, + float4 offset[3], + SMAATexture2D(predicationTex)) { + float3 neighbours = SMAAGatherNeighbours(texcoord, offset, SMAATexturePass2D(predicationTex)); + float2 delta = abs(neighbours.xx - neighbours.yz); + float2 edges = step(SMAA_PREDICATION_THRESHOLD, delta); + return SMAA_PREDICATION_SCALE * SMAA_THRESHOLD * (1.0 - SMAA_PREDICATION_STRENGTH * edges); +} + +/** + * Conditional move: + */ +void SMAAMovc(bool2 cond, inout float2 variable, float2 value) { + SMAA_FLATTEN if (cond.x) variable.x = value.x; + SMAA_FLATTEN if (cond.y) variable.y = value.y; +} + +void SMAAMovc(bool4 cond, inout float4 variable, float4 value) { + SMAAMovc(cond.xy, variable.xy, value.xy); + SMAAMovc(cond.zw, variable.zw, value.zw); +} + + +#if SMAA_INCLUDE_VS +//----------------------------------------------------------------------------- +// Vertex Shaders + +/** + * Edge Detection Vertex Shader + */ +void SMAAEdgeDetectionVS(float2 texcoord, + out float4 offset[3]) { + offset[0] = mad(SMAA_RT_METRICS.xyxy, float4(-1.0, 0.0, 0.0, -1.0), texcoord.xyxy); + offset[1] = mad(SMAA_RT_METRICS.xyxy, float4( 1.0, 0.0, 0.0, 1.0), texcoord.xyxy); + offset[2] = mad(SMAA_RT_METRICS.xyxy, float4(-2.0, 0.0, 0.0, -2.0), texcoord.xyxy); +} + +/** + * Blend Weight Calculation Vertex Shader + */ +void SMAABlendingWeightCalculationVS(float2 texcoord, + out float2 pixcoord, + out float4 offset[3]) { + pixcoord = texcoord * SMAA_RT_METRICS.zw; + + // We will use these offsets for the searches later on (see @PSEUDO_GATHER4): + offset[0] = mad(SMAA_RT_METRICS.xyxy, float4(-0.25, -0.125, 1.25, -0.125), texcoord.xyxy); + offset[1] = mad(SMAA_RT_METRICS.xyxy, float4(-0.125, -0.25, -0.125, 1.25), texcoord.xyxy); + + // And these for the searches, they indicate the ends of the loops: + offset[2] = mad(SMAA_RT_METRICS.xxyy, + float4(-2.0, 2.0, -2.0, 2.0) * float(SMAA_MAX_SEARCH_STEPS), + float4(offset[0].xz, offset[1].yw)); +} + +/** + * Neighborhood Blending Vertex Shader + */ +void SMAANeighborhoodBlendingVS(float2 texcoord, + out float4 offset) { + offset = mad(SMAA_RT_METRICS.xyxy, float4( 1.0, 0.0, 0.0, 1.0), texcoord.xyxy); +} +#endif // SMAA_INCLUDE_VS + +#if SMAA_INCLUDE_PS +//----------------------------------------------------------------------------- +// Edge Detection Pixel Shaders (First Pass) + +/** + * Luma Edge Detection + * + * IMPORTANT NOTICE: luma edge detection requires gamma-corrected colors, and + * thus 'colorTex' should be a non-sRGB texture. + */ +float2 SMAALumaEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(colorTex) + #if SMAA_PREDICATION + , SMAATexture2D(predicationTex) + #endif + ) { + // Calculate the threshold: + #if SMAA_PREDICATION + float2 threshold = SMAACalculatePredicatedThreshold(texcoord, offset, SMAATexturePass2D(predicationTex)); + #else + float2 threshold = float2(SMAA_THRESHOLD, SMAA_THRESHOLD); + #endif + + // Calculate lumas: + float3 weights = float3(0.2126, 0.7152, 0.0722); + float L = dot(SMAASamplePoint(colorTex, texcoord).rgb, weights); + + float Lleft = dot(SMAASamplePoint(colorTex, offset[0].xy).rgb, weights); + float Ltop = dot(SMAASamplePoint(colorTex, offset[0].zw).rgb, weights); + + // We do the usual threshold: + float4 delta; + delta.xy = abs(L - float2(Lleft, Ltop)); + float2 edges = step(threshold, delta.xy); + + // Then discard if there is no edge: + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + // Calculate right and bottom deltas: + float Lright = dot(SMAASamplePoint(colorTex, offset[1].xy).rgb, weights); + float Lbottom = dot(SMAASamplePoint(colorTex, offset[1].zw).rgb, weights); + delta.zw = abs(L - float2(Lright, Lbottom)); + + // Calculate the maximum delta in the direct neighborhood: + float2 maxDelta = max(delta.xy, delta.zw); + + // Calculate left-left and top-top deltas: + float Lleftleft = dot(SMAASamplePoint(colorTex, offset[2].xy).rgb, weights); + float Ltoptop = dot(SMAASamplePoint(colorTex, offset[2].zw).rgb, weights); + delta.zw = abs(float2(Lleft, Ltop) - float2(Lleftleft, Ltoptop)); + + // Calculate the final maximum delta: + maxDelta = max(maxDelta.xy, delta.zw); + float finalDelta = max(maxDelta.x, maxDelta.y); + + // Local contrast adaptation: + edges.xy *= step(finalDelta, SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR * delta.xy); + + return edges; +} + +/** + * Color Edge Detection + * + * IMPORTANT NOTICE: color edge detection requires gamma-corrected colors, and + * thus 'colorTex' should be a non-sRGB texture. + */ +float2 SMAAColorEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(colorTex) + #if SMAA_PREDICATION + , SMAATexture2D(predicationTex) + #endif + ) { + // Calculate the threshold: + #if SMAA_PREDICATION + float2 threshold = SMAACalculatePredicatedThreshold(texcoord, offset, predicationTex); + #else + float2 threshold = float2(SMAA_THRESHOLD, SMAA_THRESHOLD); + #endif + + // Calculate color deltas: + float4 delta; + float3 C = SMAASamplePoint(colorTex, texcoord).rgb; + + float3 Cleft = SMAASamplePoint(colorTex, offset[0].xy).rgb; + float3 t = abs(C - Cleft); + delta.x = max(max(t.r, t.g), t.b); + + float3 Ctop = SMAASamplePoint(colorTex, offset[0].zw).rgb; + t = abs(C - Ctop); + delta.y = max(max(t.r, t.g), t.b); + + // We do the usual threshold: + float2 edges = step(threshold, delta.xy); + + // Then discard if there is no edge: + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + // Calculate right and bottom deltas: + float3 Cright = SMAASamplePoint(colorTex, offset[1].xy).rgb; + t = abs(C - Cright); + delta.z = max(max(t.r, t.g), t.b); + + float3 Cbottom = SMAASamplePoint(colorTex, offset[1].zw).rgb; + t = abs(C - Cbottom); + delta.w = max(max(t.r, t.g), t.b); + + // Calculate the maximum delta in the direct neighborhood: + float2 maxDelta = max(delta.xy, delta.zw); + + // Calculate left-left and top-top deltas: + float3 Cleftleft = SMAASamplePoint(colorTex, offset[2].xy).rgb; + t = abs(C - Cleftleft); + delta.z = max(max(t.r, t.g), t.b); + + float3 Ctoptop = SMAASamplePoint(colorTex, offset[2].zw).rgb; + t = abs(C - Ctoptop); + delta.w = max(max(t.r, t.g), t.b); + + // Calculate the final maximum delta: + maxDelta = max(maxDelta.xy, delta.zw); + float finalDelta = max(maxDelta.x, maxDelta.y); + + // Local contrast adaptation: + edges.xy *= step(finalDelta, SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR * delta.xy); + + return edges; +} + +/** + * Depth Edge Detection + */ +float2 SMAADepthEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(depthTex)) { + float3 neighbours = SMAAGatherNeighbours(texcoord, offset, SMAATexturePass2D(depthTex)); + float2 delta = abs(neighbours.xx - float2(neighbours.y, neighbours.z)); + float2 edges = step(SMAA_DEPTH_THRESHOLD, delta); + + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + return edges; +} + +//----------------------------------------------------------------------------- +// Diagonal Search Functions + +#if !defined(SMAA_DISABLE_DIAG_DETECTION) + +/** + * Allows to decode two binary values from a bilinear-filtered access. + */ +float2 SMAADecodeDiagBilinearAccess(float2 e) { + // Bilinear access for fetching 'e' have a 0.25 offset, and we are + // interested in the R and G edges: + // + // +---G---+-------+ + // | x o R x | + // +-------+-------+ + // + // Then, if one of these edge is enabled: + // Red: (0.75 * X + 0.25 * 1) => 0.25 or 1.0 + // Green: (0.75 * 1 + 0.25 * X) => 0.75 or 1.0 + // + // This function will unpack the values (mad + mul + round): + // wolframalpha.com: round(x * abs(5 * x - 5 * 0.75)) plot 0 to 1 + e.r = e.r * abs(5.0 * e.r - 5.0 * 0.75); + return round(e); +} + +float4 SMAADecodeDiagBilinearAccess(float4 e) { + e.rb = e.rb * abs(5.0 * e.rb - 5.0 * 0.75); + return round(e); +} + +/** + * These functions allows to perform diagonal pattern searches. + */ +float2 SMAASearchDiag1(SMAATexture2D(edgesTex), float2 texcoord, float2 dir, out float2 e) { + float4 coord = float4(texcoord, -1.0, 1.0); + float3 t = float3(SMAA_RT_METRICS.xy, 1.0); + while (coord.z < float(SMAA_MAX_SEARCH_STEPS_DIAG - 1) && + coord.w > 0.9) { + coord.xyz = mad(t, float3(dir, 1.0), coord.xyz); + e = SMAASampleLevelZero(edgesTex, coord.xy).rg; + coord.w = dot(e, float2(0.5, 0.5)); + } + return coord.zw; +} + +float2 SMAASearchDiag2(SMAATexture2D(edgesTex), float2 texcoord, float2 dir, out float2 e) { + float4 coord = float4(texcoord, -1.0, 1.0); + coord.x += 0.25 * SMAA_RT_METRICS.x; // See @SearchDiag2Optimization + float3 t = float3(SMAA_RT_METRICS.xy, 1.0); + while (coord.z < float(SMAA_MAX_SEARCH_STEPS_DIAG - 1) && + coord.w > 0.9) { + coord.xyz = mad(t, float3(dir, 1.0), coord.xyz); + + // @SearchDiag2Optimization + // Fetch both edges at once using bilinear filtering: + e = SMAASampleLevelZero(edgesTex, coord.xy).rg; + e = SMAADecodeDiagBilinearAccess(e); + + // Non-optimized version: + // e.g = SMAASampleLevelZero(edgesTex, coord.xy).g; + // e.r = SMAASampleLevelZeroOffset(edgesTex, coord.xy, int2(1, 0)).r; + + coord.w = dot(e, float2(0.5, 0.5)); + } + return coord.zw; +} + +/** + * Similar to SMAAArea, this calculates the area corresponding to a certain + * diagonal distance and crossing edges 'e'. + */ +float2 SMAAAreaDiag(SMAATexture2D(areaTex), float2 dist, float2 e, float offset) { + float2 texcoord = mad(float2(SMAA_AREATEX_MAX_DISTANCE_DIAG, SMAA_AREATEX_MAX_DISTANCE_DIAG), e, dist); + + // We do a scale and bias for mapping to texel space: + texcoord = mad(SMAA_AREATEX_PIXEL_SIZE, texcoord, 0.5 * SMAA_AREATEX_PIXEL_SIZE); + + // Diagonal areas are on the second half of the texture: + texcoord.x += 0.5; + + // Move to proper place, according to the subpixel offset: + texcoord.y += SMAA_AREATEX_SUBTEX_SIZE * offset; + + // Do it! + return SMAA_AREATEX_SELECT(SMAASampleLevelZero(areaTex, texcoord)); +} + +/** + * This searches for diagonal patterns and returns the corresponding weights. + */ +float2 SMAACalculateDiagWeights(SMAATexture2D(edgesTex), SMAATexture2D(areaTex), float2 texcoord, float2 e, float4 subsampleIndices) { + float2 weights = float2(0.0, 0.0); + + // Search for the line ends: + float4 d; + float2 end; + if (e.r > 0.0) { + d.xz = SMAASearchDiag1(SMAATexturePass2D(edgesTex), texcoord, float2(-1.0, 1.0), end); + d.x += float(end.y > 0.9); + } else + d.xz = float2(0.0, 0.0); + d.yw = SMAASearchDiag1(SMAATexturePass2D(edgesTex), texcoord, float2(1.0, -1.0), end); + + SMAA_BRANCH + if (d.x + d.y > 2.0) { // d.x + d.y + 1 > 3 + // Fetch the crossing edges: + float4 coords = mad(float4(-d.x + 0.25, d.x, d.y, -d.y - 0.25), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + float4 c; + c.xy = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).rg; + c.zw = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).rg; + c.yxwz = SMAADecodeDiagBilinearAccess(c.xyzw); + + // Non-optimized version: + // float4 coords = mad(float4(-d.x, d.x, d.y, -d.y), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + // float4 c; + // c.x = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).g; + // c.y = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2( 0, 0)).r; + // c.z = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).g; + // c.w = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, -1)).r; + + // Merge crossing edges at each side into a single value: + float2 cc = mad(float2(2.0, 2.0), c.xz, c.yw); + + // Remove the crossing edge if we didn't found the end of the line: + SMAAMovc(bool2(step(0.9, d.zw)), cc, float2(0.0, 0.0)); + + // Fetch the areas for this line: + weights += SMAAAreaDiag(SMAATexturePass2D(areaTex), d.xy, cc, subsampleIndices.z); + } + + // Search for the line ends: + d.xz = SMAASearchDiag2(SMAATexturePass2D(edgesTex), texcoord, float2(-1.0, -1.0), end); + if (SMAASampleLevelZeroOffset(edgesTex, texcoord, int2(1, 0)).r > 0.0) { + d.yw = SMAASearchDiag2(SMAATexturePass2D(edgesTex), texcoord, float2(1.0, 1.0), end); + d.y += float(end.y > 0.9); + } else + d.yw = float2(0.0, 0.0); + + SMAA_BRANCH + if (d.x + d.y > 2.0) { // d.x + d.y + 1 > 3 + // Fetch the crossing edges: + float4 coords = mad(float4(-d.x, -d.x, d.y, d.y), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + float4 c; + c.x = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).g; + c.y = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2( 0, -1)).r; + c.zw = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).gr; + float2 cc = mad(float2(2.0, 2.0), c.xz, c.yw); + + // Remove the crossing edge if we didn't found the end of the line: + SMAAMovc(bool2(step(0.9, d.zw)), cc, float2(0.0, 0.0)); + + // Fetch the areas for this line: + weights += SMAAAreaDiag(SMAATexturePass2D(areaTex), d.xy, cc, subsampleIndices.w).gr; + } + + return weights; +} +#endif + +//----------------------------------------------------------------------------- +// Horizontal/Vertical Search Functions + +/** + * This allows to determine how much length should we add in the last step + * of the searches. It takes the bilinearly interpolated edge (see + * @PSEUDO_GATHER4), and adds 0, 1 or 2, depending on which edges and + * crossing edges are active. + */ +float SMAASearchLength(SMAATexture2D(searchTex), float2 e, float offset) { + // The texture is flipped vertically, with left and right cases taking half + // of the space horizontally: + float2 scale = SMAA_SEARCHTEX_SIZE * float2(0.5, -1.0); + float2 bias = SMAA_SEARCHTEX_SIZE * float2(offset, 1.0); + + // Scale and bias to access texel centers: + scale += float2(-1.0, 1.0); + bias += float2( 0.5, -0.5); + + // Convert from pixel coordinates to texcoords: + // (We use SMAA_SEARCHTEX_PACKED_SIZE because the texture is cropped) + scale *= 1.0 / SMAA_SEARCHTEX_PACKED_SIZE; + bias *= 1.0 / SMAA_SEARCHTEX_PACKED_SIZE; + + // Lookup the search texture: + return SMAA_SEARCHTEX_SELECT(SMAASampleLevelZero(searchTex, mad(scale, e, bias))); +} + +/** + * Horizontal/vertical search functions for the 2nd pass. + */ +float SMAASearchXLeft(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + /** + * @PSEUDO_GATHER4 + * This texcoord has been offset by (-0.25, -0.125) in the vertex shader to + * sample between edge, thus fetching four edges in a row. + * Sampling with different offsets in each direction allows to disambiguate + * which edges are active from the four fetched ones. + */ + float2 e = float2(0.0, 1.0); + while (texcoord.x > end && + e.g > 0.8281 && // Is there some edge not activated? + e.r == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(-float2(2.0, 0.0), SMAA_RT_METRICS.xy, texcoord); + } + + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.0), 3.25); + return mad(SMAA_RT_METRICS.x, offset, texcoord.x); + + // Non-optimized version: + // We correct the previous (-0.25, -0.125) offset we applied: + // texcoord.x += 0.25 * SMAA_RT_METRICS.x; + + // The searches are bias by 1, so adjust the coords accordingly: + // texcoord.x += SMAA_RT_METRICS.x; + + // Disambiguate the length added by the last step: + // texcoord.x += 2.0 * SMAA_RT_METRICS.x; // Undo last step + // texcoord.x -= SMAA_RT_METRICS.x * (255.0 / 127.0) * SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.0); + // return mad(SMAA_RT_METRICS.x, offset, texcoord.x); +} + +float SMAASearchXRight(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(0.0, 1.0); + while (texcoord.x < end && + e.g > 0.8281 && // Is there some edge not activated? + e.r == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(float2(2.0, 0.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.5), 3.25); + return mad(-SMAA_RT_METRICS.x, offset, texcoord.x); +} + +float SMAASearchYUp(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(1.0, 0.0); + while (texcoord.y > end && + e.r > 0.8281 && // Is there some edge not activated? + e.g == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(-float2(0.0, 2.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e.gr, 0.0), 3.25); + return mad(SMAA_RT_METRICS.y, offset, texcoord.y); +} + +float SMAASearchYDown(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(1.0, 0.0); + while (texcoord.y < end && + e.r > 0.8281 && // Is there some edge not activated? + e.g == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(float2(0.0, 2.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e.gr, 0.5), 3.25); + return mad(-SMAA_RT_METRICS.y, offset, texcoord.y); +} + +/** + * Ok, we have the distance and both crossing edges. So, what are the areas + * at each side of current edge? + */ +float2 SMAAArea(SMAATexture2D(areaTex), float2 dist, float e1, float e2, float offset) { + // Rounding prevents precision errors of bilinear filtering: + float2 texcoord = mad(float2(SMAA_AREATEX_MAX_DISTANCE, SMAA_AREATEX_MAX_DISTANCE), round(4.0 * float2(e1, e2)), dist); + + // We do a scale and bias for mapping to texel space: + texcoord = mad(SMAA_AREATEX_PIXEL_SIZE, texcoord, 0.5 * SMAA_AREATEX_PIXEL_SIZE); + + // Move to proper place, according to the subpixel offset: + texcoord.y = mad(SMAA_AREATEX_SUBTEX_SIZE, offset, texcoord.y); + + // Do it! + return SMAA_AREATEX_SELECT(SMAASampleLevelZero(areaTex, texcoord)); +} + +//----------------------------------------------------------------------------- +// Corner Detection Functions + +void SMAADetectHorizontalCornerPattern(SMAATexture2D(edgesTex), inout float2 weights, float4 texcoord, float2 d) { + #if !defined(SMAA_DISABLE_CORNER_DETECTION) + float2 leftRight = step(d.xy, d.yx); + float2 rounding = (1.0 - SMAA_CORNER_ROUNDING_NORM) * leftRight; + + rounding /= leftRight.x + leftRight.y; // Reduce blending for pixels in the center of a line. + + float2 factor = float2(1.0, 1.0); + factor.x -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(0, 1)).r; + factor.x -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(1, 1)).r; + factor.y -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(0, -2)).r; + factor.y -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(1, -2)).r; + + weights *= saturate(factor); + #endif +} + +void SMAADetectVerticalCornerPattern(SMAATexture2D(edgesTex), inout float2 weights, float4 texcoord, float2 d) { + #if !defined(SMAA_DISABLE_CORNER_DETECTION) + float2 leftRight = step(d.xy, d.yx); + float2 rounding = (1.0 - SMAA_CORNER_ROUNDING_NORM) * leftRight; + + rounding /= leftRight.x + leftRight.y; + + float2 factor = float2(1.0, 1.0); + factor.x -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2( 1, 0)).g; + factor.x -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2( 1, 1)).g; + factor.y -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(-2, 0)).g; + factor.y -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(-2, 1)).g; + + weights *= saturate(factor); + #endif +} + +//----------------------------------------------------------------------------- +// Blending Weight Calculation Pixel Shader (Second Pass) + +float4 SMAABlendingWeightCalculationPS(float2 texcoord, + float2 pixcoord, + float4 offset[3], + SMAATexture2D(edgesTex), + SMAATexture2D(areaTex), + SMAATexture2D(searchTex), + float4 subsampleIndices) { // Just pass zero for SMAA 1x, see @SUBSAMPLE_INDICES. + float4 weights = float4(0.0, 0.0, 0.0, 0.0); + + float2 e = SMAASample(edgesTex, texcoord).rg; + + SMAA_BRANCH + if (e.g > 0.0) { // Edge at north + #if !defined(SMAA_DISABLE_DIAG_DETECTION) + // Diagonals have both north and west edges, so searching for them in + // one of the boundaries is enough. + weights.rg = SMAACalculateDiagWeights(SMAATexturePass2D(edgesTex), SMAATexturePass2D(areaTex), texcoord, e, subsampleIndices); + + // We give priority to diagonals, so if we find a diagonal we skip + // horizontal/vertical processing. + SMAA_BRANCH + if (weights.r == -weights.g) { // weights.r + weights.g == 0.0 + #endif + + float2 d; + + // Find the distance to the left: + float3 coords; + coords.x = SMAASearchXLeft(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[0].xy, offset[2].x); + coords.y = offset[1].y; // offset[1].y = texcoord.y - 0.25 * SMAA_RT_METRICS.y (@CROSSING_OFFSET) + d.x = coords.x; + + // Now fetch the left crossing edges, two at a time using bilinear + // filtering. Sampling at -0.25 (see @CROSSING_OFFSET) enables to + // discern what value each edge has: + float e1 = SMAASampleLevelZero(edgesTex, coords.xy).r; + + // Find the distance to the right: + coords.z = SMAASearchXRight(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[0].zw, offset[2].y); + d.y = coords.z; + + // We want the distances to be in pixel units (doing this here allow to + // better interleave arithmetic and memory accesses): + d = abs(round(mad(SMAA_RT_METRICS.zz, d, -pixcoord.xx))); + + // SMAAArea below needs a sqrt, as the areas texture is compressed + // quadratically: + float2 sqrt_d = sqrt(d); + + // Fetch the right crossing edges: + float e2 = SMAASampleLevelZeroOffset(edgesTex, coords.zy, int2(1, 0)).r; + + // Ok, we know how this pattern looks like, now it is time for getting + // the actual area: + weights.rg = SMAAArea(SMAATexturePass2D(areaTex), sqrt_d, e1, e2, subsampleIndices.y); + + // Fix corners: + coords.y = texcoord.y; + SMAADetectHorizontalCornerPattern(SMAATexturePass2D(edgesTex), weights.rg, coords.xyzy, d); + + #if !defined(SMAA_DISABLE_DIAG_DETECTION) + } else + e.r = 0.0; // Skip vertical processing. + #endif + } + + SMAA_BRANCH + if (e.r > 0.0) { // Edge at west + float2 d; + + // Find the distance to the top: + float3 coords; + coords.y = SMAASearchYUp(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[1].xy, offset[2].z); + coords.x = offset[0].x; // offset[1].x = texcoord.x - 0.25 * SMAA_RT_METRICS.x; + d.x = coords.y; + + // Fetch the top crossing edges: + float e1 = SMAASampleLevelZero(edgesTex, coords.xy).g; + + // Find the distance to the bottom: + coords.z = SMAASearchYDown(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[1].zw, offset[2].w); + d.y = coords.z; + + // We want the distances to be in pixel units: + d = abs(round(mad(SMAA_RT_METRICS.ww, d, -pixcoord.yy))); + + // SMAAArea below needs a sqrt, as the areas texture is compressed + // quadratically: + float2 sqrt_d = sqrt(d); + + // Fetch the bottom crossing edges: + float e2 = SMAASampleLevelZeroOffset(edgesTex, coords.xz, int2(0, 1)).g; + + // Get the area for this direction: + weights.ba = SMAAArea(SMAATexturePass2D(areaTex), sqrt_d, e1, e2, subsampleIndices.x); + + // Fix corners: + coords.x = texcoord.x; + SMAADetectVerticalCornerPattern(SMAATexturePass2D(edgesTex), weights.ba, coords.xyxz, d); + } + + return weights; +} + +//----------------------------------------------------------------------------- +// Neighborhood Blending Pixel Shader (Third Pass) + +float4 SMAANeighborhoodBlendingPS(float2 texcoord, + float4 offset, + SMAATexture2D(colorTex), + SMAATexture2D(blendTex) + #if SMAA_REPROJECTION + , SMAATexture2D(velocityTex) + #endif + ) { + // Fetch the blending weights for current pixel: + float4 a; + a.x = SMAASample(blendTex, offset.xy).a; // Right + a.y = SMAASample(blendTex, offset.zw).g; // Top + a.wz = SMAASample(blendTex, texcoord).xz; // Bottom / Left + + // Is there any blending weight with a value greater than 0.0? + SMAA_BRANCH + if (dot(a, float4(1.0, 1.0, 1.0, 1.0)) < 1e-5) { + float4 color = SMAASampleLevelZero(colorTex, texcoord); + + #if SMAA_REPROJECTION + float2 velocity = SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, texcoord)); + + // Pack velocity into the alpha channel: + color.a = sqrt(5.0 * length(velocity)); + #endif + + return color; + } else { + bool h = max(a.x, a.z) > max(a.y, a.w); // max(horizontal) > max(vertical) + + // Calculate the blending offsets: + float4 blendingOffset = float4(0.0, a.y, 0.0, a.w); + float2 blendingWeight = a.yw; + SMAAMovc(bool4(h, h, h, h), blendingOffset, float4(a.x, 0.0, a.z, 0.0)); + SMAAMovc(bool2(h, h), blendingWeight, a.xz); + blendingWeight /= dot(blendingWeight, float2(1.0, 1.0)); + + // Calculate the texture coordinates: + float4 blendingCoord = mad(blendingOffset, float4(SMAA_RT_METRICS.xy, -SMAA_RT_METRICS.xy), texcoord.xyxy); + + // We exploit bilinear filtering to mix current pixel with the chosen + // neighbor: + float4 color = blendingWeight.x * SMAASampleLevelZero(colorTex, blendingCoord.xy); + color += blendingWeight.y * SMAASampleLevelZero(colorTex, blendingCoord.zw); + + #if SMAA_REPROJECTION + // Antialias velocity for proper reprojection in a later stage: + float2 velocity = blendingWeight.x * SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, blendingCoord.xy)); + velocity += blendingWeight.y * SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, blendingCoord.zw)); + + // Pack velocity into the alpha channel: + color.a = sqrt(5.0 * length(velocity)); + #endif + + return color; + } +} + +//----------------------------------------------------------------------------- +// Temporal Resolve Pixel Shader (Optional Pass) + +float4 SMAAResolvePS(float2 texcoord, + SMAATexture2D(currentColorTex), + SMAATexture2D(previousColorTex) + #if SMAA_REPROJECTION + , SMAATexture2D(velocityTex) + #endif + ) { + #if SMAA_REPROJECTION + // Velocity is assumed to be calculated for motion blur, so we need to + // inverse it for reprojection: + float2 velocity = -SMAA_DECODE_VELOCITY(SMAASamplePoint(velocityTex, texcoord).rg); + + // Fetch current pixel: + float4 current = SMAASamplePoint(currentColorTex, texcoord); + + // Reproject current coordinates and fetch previous pixel: + float4 previous = SMAASamplePoint(previousColorTex, texcoord + velocity); + + // Attenuate the previous pixel if the velocity is different: + float delta = abs(current.a * current.a - previous.a * previous.a) / 5.0; + float weight = 0.5 * saturate(1.0 - sqrt(delta) * SMAA_REPROJECTION_WEIGHT_SCALE); + + // Blend the pixels according to the calculated weight: + return lerp(current, previous, weight); + #else + // Just blend the pixels: + float4 current = SMAASamplePoint(currentColorTex, texcoord); + float4 previous = SMAASamplePoint(previousColorTex, texcoord); + return lerp(current, previous, 0.5); + #endif +} + +//----------------------------------------------------------------------------- +// Separate Multisamples Pixel Shader (Optional Pass) + +#ifdef SMAALoad +void SMAASeparatePS(float4 position, + float2 texcoord, + out float4 target0, + out float4 target1, + SMAATexture2DMS2(colorTexMS)) { + int2 pos = int2(position.xy); + target0 = SMAALoad(colorTexMS, pos, 0); + target1 = SMAALoad(colorTexMS, pos, 1); +} +#endif + +//----------------------------------------------------------------------------- +#endif // SMAA_INCLUDE_PS + +layout(rgba8, binding = 0, set = 3) uniform image2D imgOutput; + +layout(binding = 1, set = 2) uniform sampler2D inputImg; +layout(binding = 3, set = 2) uniform sampler2D samplerArea; +layout(binding = 4, set = 2) uniform sampler2D samplerSearch; +layout( binding = 2 ) uniform invResolution +{ + vec2 invResolution_data; +}; + +void main() { + ivec2 loc = ivec2(gl_GlobalInvocationID.x * 4, gl_GlobalInvocationID.y * 4); + for(int i = 0; i < 4; i++) + { + for(int j = 0; j < 4; j++) + { + ivec2 texelCoord = ivec2(loc.x + i, loc.y + j); + vec2 coord = (texelCoord + vec2(0.5)) / invResolution_data; + vec2 pixCoord; + vec4 offset[3]; + + SMAABlendingWeightCalculationVS( coord, pixCoord, offset); + + vec4 oColor = SMAABlendingWeightCalculationPS(coord, pixCoord, offset, inputImg, samplerArea, samplerSearch, ivec4(0)); + + imageStore(imgOutput, texelCoord, oColor); + } + } +} diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaBlend.spv b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaBlend.spv new file mode 100644 index 00000000..8efa011f Binary files /dev/null and b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaBlend.spv differ diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaEdge.glsl b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaEdge.glsl new file mode 100644 index 00000000..668b97d5 --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaEdge.glsl @@ -0,0 +1,1402 @@ +#version 430 core +#define SMAA_GLSL_4 1 + +layout (constant_id = 0) const int SMAA_PRESET_LOW = 0; +layout (constant_id = 1) const int SMAA_PRESET_MEDIUM = 0; +layout (constant_id = 2) const int SMAA_PRESET_HIGH = 0; +layout (constant_id = 3) const int SMAA_PRESET_ULTRA = 0; +layout (constant_id = 4) const float METRIC_WIDTH = 1920.0; +layout (constant_id = 5) const float METRIC_HEIGHT = 1080.0; + +#define SMAA_RT_METRICS float4(1.0 / METRIC_WIDTH, 1.0 / METRIC_HEIGHT, METRIC_WIDTH, METRIC_HEIGHT) + +layout (local_size_x = 16, local_size_y = 16) in; +/** + * Copyright (C) 2013 Jorge Jimenez (jorge@iryoku.com) + * Copyright (C) 2013 Jose I. Echevarria (joseignacioechevarria@gmail.com) + * Copyright (C) 2013 Belen Masia (bmasia@unizar.es) + * Copyright (C) 2013 Fernando Navarro (fernandn@microsoft.com) + * Copyright (C) 2013 Diego Gutierrez (diegog@unizar.es) + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * this software and associated documentation files (the "Software"), to deal in + * the Software without restriction, including without limitation the rights to + * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies + * of the Software, and to permit persons to whom the Software is furnished to + * do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. As clarification, there + * is no requirement that the copyright notice and permission be included in + * binary distributions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + + +/** + * _______ ___ ___ ___ ___ + * / || \/ | / \ / \ + * | (---- | \ / | / ^ \ / ^ \ + * \ \ | |\/| | / /_\ \ / /_\ \ + * ----) | | | | | / _____ \ / _____ \ + * |_______/ |__| |__| /__/ \__\ /__/ \__\ + * + * E N H A N C E D + * S U B P I X E L M O R P H O L O G I C A L A N T I A L I A S I N G + * + * http://www.iryoku.com/smaa/ + * + * Hi, welcome aboard! + * + * Here you'll find instructions to get the shader up and running as fast as + * possible. + * + * IMPORTANTE NOTICE: when updating, remember to update both this file and the + * precomputed textures! They may change from version to version. + * + * The shader has three passes, chained together as follows: + * + * |input|------------------ + * v | + * [ SMAA*EdgeDetection ] | + * v | + * |edgesTex| | + * v | + * [ SMAABlendingWeightCalculation ] | + * v | + * |blendTex| | + * v | + * [ SMAANeighborhoodBlending ] <------ + * v + * |output| + * + * Note that each [pass] has its own vertex and pixel shader. Remember to use + * oversized triangles instead of quads to avoid overshading along the + * diagonal. + * + * You've three edge detection methods to choose from: luma, color or depth. + * They represent different quality/performance and anti-aliasing/sharpness + * tradeoffs, so our recommendation is for you to choose the one that best + * suits your particular scenario: + * + * - Depth edge detection is usually the fastest but it may miss some edges. + * + * - Luma edge detection is usually more expensive than depth edge detection, + * but catches visible edges that depth edge detection can miss. + * + * - Color edge detection is usually the most expensive one but catches + * chroma-only edges. + * + * For quickstarters: just use luma edge detection. + * + * The general advice is to not rush the integration process and ensure each + * step is done correctly (don't try to integrate SMAA T2x with predicated edge + * detection from the start!). Ok then, let's go! + * + * 1. The first step is to create two RGBA temporal render targets for holding + * |edgesTex| and |blendTex|. + * + * In DX10 or DX11, you can use a RG render target for the edges texture. + * In the case of NVIDIA GPUs, using RG render targets seems to actually be + * slower. + * + * On the Xbox 360, you can use the same render target for resolving both + * |edgesTex| and |blendTex|, as they aren't needed simultaneously. + * + * 2. Both temporal render targets |edgesTex| and |blendTex| must be cleared + * each frame. Do not forget to clear the alpha channel! + * + * 3. The next step is loading the two supporting precalculated textures, + * 'areaTex' and 'searchTex'. You'll find them in the 'Textures' folder as + * C++ headers, and also as regular DDS files. They'll be needed for the + * 'SMAABlendingWeightCalculation' pass. + * + * If you use the C++ headers, be sure to load them in the format specified + * inside of them. + * + * You can also compress 'areaTex' and 'searchTex' using BC5 and BC4 + * respectively, if you have that option in your content processor pipeline. + * When compressing then, you get a non-perceptible quality decrease, and a + * marginal performance increase. + * + * 4. All samplers must be set to linear filtering and clamp. + * + * After you get the technique working, remember that 64-bit inputs have + * half-rate linear filtering on GCN. + * + * If SMAA is applied to 64-bit color buffers, switching to point filtering + * when accesing them will increase the performance. Search for + * 'SMAASamplePoint' to see which textures may benefit from point + * filtering, and where (which is basically the color input in the edge + * detection and resolve passes). + * + * 5. All texture reads and buffer writes must be non-sRGB, with the exception + * of the input read and the output write in + * 'SMAANeighborhoodBlending' (and only in this pass!). If sRGB reads in + * this last pass are not possible, the technique will work anyway, but + * will perform antialiasing in gamma space. + * + * IMPORTANT: for best results the input read for the color/luma edge + * detection should *NOT* be sRGB. + * + * 6. Before including SMAA.h you'll have to setup the render target metrics, + * the target and any optional configuration defines. Optionally you can + * use a preset. + * + * You have the following targets available: + * SMAA_HLSL_3 + * SMAA_HLSL_4 + * SMAA_HLSL_4_1 + * SMAA_GLSL_3 * + * SMAA_GLSL_4 * + * + * * (See SMAA_INCLUDE_VS and SMAA_INCLUDE_PS below). + * + * And four presets: + * SMAA_PRESET_LOW (%60 of the quality) + * SMAA_PRESET_MEDIUM (%80 of the quality) + * SMAA_PRESET_HIGH (%95 of the quality) + * SMAA_PRESET_ULTRA (%99 of the quality) + * + * For example: + * #define SMAA_RT_METRICS float4(1.0 / 1280.0, 1.0 / 720.0, 1280.0, 720.0) + * #define SMAA_HLSL_4 + * #define SMAA_PRESET_HIGH + * #include "SMAA.h" + * + * Note that SMAA_RT_METRICS doesn't need to be a macro, it can be a + * uniform variable. The code is designed to minimize the impact of not + * using a constant value, but it is still better to hardcode it. + * + * Depending on how you encoded 'areaTex' and 'searchTex', you may have to + * add (and customize) the following defines before including SMAA.h: + * #define SMAA_AREATEX_SELECT(sample) sample.rg + * #define SMAA_SEARCHTEX_SELECT(sample) sample.r + * + * If your engine is already using porting macros, you can define + * SMAA_CUSTOM_SL, and define the porting functions by yourself. + * + * 7. Then, you'll have to setup the passes as indicated in the scheme above. + * You can take a look into SMAA.fx, to see how we did it for our demo. + * Checkout the function wrappers, you may want to copy-paste them! + * + * 8. It's recommended to validate the produced |edgesTex| and |blendTex|. + * You can use a screenshot from your engine to compare the |edgesTex| + * and |blendTex| produced inside of the engine with the results obtained + * with the reference demo. + * + * 9. After you get the last pass to work, it's time to optimize. You'll have + * to initialize a stencil buffer in the first pass (discard is already in + * the code), then mask execution by using it the second pass. The last + * pass should be executed in all pixels. + * + * + * After this point you can choose to enable predicated thresholding, + * temporal supersampling and motion blur integration: + * + * a) If you want to use predicated thresholding, take a look into + * SMAA_PREDICATION; you'll need to pass an extra texture in the edge + * detection pass. + * + * b) If you want to enable temporal supersampling (SMAA T2x): + * + * 1. The first step is to render using subpixel jitters. I won't go into + * detail, but it's as simple as moving each vertex position in the + * vertex shader, you can check how we do it in our DX10 demo. + * + * 2. Then, you must setup the temporal resolve. You may want to take a look + * into SMAAResolve for resolving 2x modes. After you get it working, you'll + * probably see ghosting everywhere. But fear not, you can enable the + * CryENGINE temporal reprojection by setting the SMAA_REPROJECTION macro. + * Check out SMAA_DECODE_VELOCITY if your velocity buffer is encoded. + * + * 3. The next step is to apply SMAA to each subpixel jittered frame, just as + * done for 1x. + * + * 4. At this point you should already have something usable, but for best + * results the proper area textures must be set depending on current jitter. + * For this, the parameter 'subsampleIndices' of + * 'SMAABlendingWeightCalculationPS' must be set as follows, for our T2x + * mode: + * + * @SUBSAMPLE_INDICES + * + * | S# | Camera Jitter | subsampleIndices | + * +----+------------------+---------------------+ + * | 0 | ( 0.25, -0.25) | float4(1, 1, 1, 0) | + * | 1 | (-0.25, 0.25) | float4(2, 2, 2, 0) | + * + * These jitter positions assume a bottom-to-top y axis. S# stands for the + * sample number. + * + * More information about temporal supersampling here: + * http://iryoku.com/aacourse/downloads/13-Anti-Aliasing-Methods-in-CryENGINE-3.pdf + * + * c) If you want to enable spatial multisampling (SMAA S2x): + * + * 1. The scene must be rendered using MSAA 2x. The MSAA 2x buffer must be + * created with: + * - DX10: see below (*) + * - DX10.1: D3D10_STANDARD_MULTISAMPLE_PATTERN or + * - DX11: D3D11_STANDARD_MULTISAMPLE_PATTERN + * + * This allows to ensure that the subsample order matches the table in + * @SUBSAMPLE_INDICES. + * + * (*) In the case of DX10, we refer the reader to: + * - SMAA::detectMSAAOrder and + * - SMAA::msaaReorder + * + * These functions allow to match the standard multisample patterns by + * detecting the subsample order for a specific GPU, and reordering + * them appropriately. + * + * 2. A shader must be run to output each subsample into a separate buffer + * (DX10 is required). You can use SMAASeparate for this purpose, or just do + * it in an existing pass (for example, in the tone mapping pass, which has + * the advantage of feeding tone mapped subsamples to SMAA, which will yield + * better results). + * + * 3. The full SMAA 1x pipeline must be run for each separated buffer, storing + * the results in the final buffer. The second run should alpha blend with + * the existing final buffer using a blending factor of 0.5. + * 'subsampleIndices' must be adjusted as in the SMAA T2x case (see point + * b). + * + * d) If you want to enable temporal supersampling on top of SMAA S2x + * (which actually is SMAA 4x): + * + * 1. SMAA 4x consists on temporally jittering SMAA S2x, so the first step is + * to calculate SMAA S2x for current frame. In this case, 'subsampleIndices' + * must be set as follows: + * + * | F# | S# | Camera Jitter | Net Jitter | subsampleIndices | + * +----+----+--------------------+-------------------+----------------------+ + * | 0 | 0 | ( 0.125, 0.125) | ( 0.375, -0.125) | float4(5, 3, 1, 3) | + * | 0 | 1 | ( 0.125, 0.125) | (-0.125, 0.375) | float4(4, 6, 2, 3) | + * +----+----+--------------------+-------------------+----------------------+ + * | 1 | 2 | (-0.125, -0.125) | ( 0.125, -0.375) | float4(3, 5, 1, 4) | + * | 1 | 3 | (-0.125, -0.125) | (-0.375, 0.125) | float4(6, 4, 2, 4) | + * + * These jitter positions assume a bottom-to-top y axis. F# stands for the + * frame number. S# stands for the sample number. + * + * 2. After calculating SMAA S2x for current frame (with the new subsample + * indices), previous frame must be reprojected as in SMAA T2x mode (see + * point b). + * + * e) If motion blur is used, you may want to do the edge detection pass + * together with motion blur. This has two advantages: + * + * 1. Pixels under heavy motion can be omitted from the edge detection process. + * For these pixels we can just store "no edge", as motion blur will take + * care of them. + * 2. The center pixel tap is reused. + * + * Note that in this case depth testing should be used instead of stenciling, + * as we have to write all the pixels in the motion blur pass. + * + * That's it! + */ + +//----------------------------------------------------------------------------- +// SMAA Presets + +/** + * Note that if you use one of these presets, the following configuration + * macros will be ignored if set in the "Configurable Defines" section. + */ + +#if defined(SMAA_PRESET_LOW) +#define SMAA_THRESHOLD 0.15 +#define SMAA_MAX_SEARCH_STEPS 4 +#define SMAA_DISABLE_DIAG_DETECTION +#define SMAA_DISABLE_CORNER_DETECTION +#elif defined(SMAA_PRESET_MEDIUM) +#define SMAA_THRESHOLD 0.1 +#define SMAA_MAX_SEARCH_STEPS 8 +#define SMAA_DISABLE_DIAG_DETECTION +#define SMAA_DISABLE_CORNER_DETECTION +#elif defined(SMAA_PRESET_HIGH) +#define SMAA_THRESHOLD 0.1 +#define SMAA_MAX_SEARCH_STEPS 16 +#define SMAA_MAX_SEARCH_STEPS_DIAG 8 +#define SMAA_CORNER_ROUNDING 25 +#elif defined(SMAA_PRESET_ULTRA) +#define SMAA_THRESHOLD 0.05 +#define SMAA_MAX_SEARCH_STEPS 32 +#define SMAA_MAX_SEARCH_STEPS_DIAG 16 +#define SMAA_CORNER_ROUNDING 25 +#endif + +//----------------------------------------------------------------------------- +// Configurable Defines + +/** + * SMAA_THRESHOLD specifies the threshold or sensitivity to edges. + * Lowering this value you will be able to detect more edges at the expense of + * performance. + * + * Range: [0, 0.5] + * 0.1 is a reasonable value, and allows to catch most visible edges. + * 0.05 is a rather overkill value, that allows to catch 'em all. + * + * If temporal supersampling is used, 0.2 could be a reasonable value, as low + * contrast edges are properly filtered by just 2x. + */ +#ifndef SMAA_THRESHOLD +#define SMAA_THRESHOLD 0.1 +#endif + +/** + * SMAA_DEPTH_THRESHOLD specifies the threshold for depth edge detection. + * + * Range: depends on the depth range of the scene. + */ +#ifndef SMAA_DEPTH_THRESHOLD +#define SMAA_DEPTH_THRESHOLD (0.1 * SMAA_THRESHOLD) +#endif + +/** + * SMAA_MAX_SEARCH_STEPS specifies the maximum steps performed in the + * horizontal/vertical pattern searches, at each side of the pixel. + * + * In number of pixels, it's actually the double. So the maximum line length + * perfectly handled by, for example 16, is 64 (by perfectly, we meant that + * longer lines won't look as good, but still antialiased). + * + * Range: [0, 112] + */ +#ifndef SMAA_MAX_SEARCH_STEPS +#define SMAA_MAX_SEARCH_STEPS 16 +#endif + +/** + * SMAA_MAX_SEARCH_STEPS_DIAG specifies the maximum steps performed in the + * diagonal pattern searches, at each side of the pixel. In this case we jump + * one pixel at time, instead of two. + * + * Range: [0, 20] + * + * On high-end machines it is cheap (between a 0.8x and 0.9x slower for 16 + * steps), but it can have a significant impact on older machines. + * + * Define SMAA_DISABLE_DIAG_DETECTION to disable diagonal processing. + */ +#ifndef SMAA_MAX_SEARCH_STEPS_DIAG +#define SMAA_MAX_SEARCH_STEPS_DIAG 8 +#endif + +/** + * SMAA_CORNER_ROUNDING specifies how much sharp corners will be rounded. + * + * Range: [0, 100] + * + * Define SMAA_DISABLE_CORNER_DETECTION to disable corner processing. + */ +#ifndef SMAA_CORNER_ROUNDING +#define SMAA_CORNER_ROUNDING 25 +#endif + +/** + * If there is an neighbor edge that has SMAA_LOCAL_CONTRAST_FACTOR times + * bigger contrast than current edge, current edge will be discarded. + * + * This allows to eliminate spurious crossing edges, and is based on the fact + * that, if there is too much contrast in a direction, that will hide + * perceptually contrast in the other neighbors. + */ +#ifndef SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR +#define SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR 2.0 +#endif + +/** + * Predicated thresholding allows to better preserve texture details and to + * improve performance, by decreasing the number of detected edges using an + * additional buffer like the light accumulation buffer, object ids or even the + * depth buffer (the depth buffer usage may be limited to indoor or short range + * scenes). + * + * It locally decreases the luma or color threshold if an edge is found in an + * additional buffer (so the global threshold can be higher). + * + * This method was developed by Playstation EDGE MLAA team, and used in + * Killzone 3, by using the light accumulation buffer. More information here: + * http://iryoku.com/aacourse/downloads/06-MLAA-on-PS3.pptx + */ +#ifndef SMAA_PREDICATION +#define SMAA_PREDICATION 0 +#endif + +/** + * Threshold to be used in the additional predication buffer. + * + * Range: depends on the input, so you'll have to find the magic number that + * works for you. + */ +#ifndef SMAA_PREDICATION_THRESHOLD +#define SMAA_PREDICATION_THRESHOLD 0.01 +#endif + +/** + * How much to scale the global threshold used for luma or color edge + * detection when using predication. + * + * Range: [1, 5] + */ +#ifndef SMAA_PREDICATION_SCALE +#define SMAA_PREDICATION_SCALE 2.0 +#endif + +/** + * How much to locally decrease the threshold. + * + * Range: [0, 1] + */ +#ifndef SMAA_PREDICATION_STRENGTH +#define SMAA_PREDICATION_STRENGTH 0.4 +#endif + +/** + * Temporal reprojection allows to remove ghosting artifacts when using + * temporal supersampling. We use the CryEngine 3 method which also introduces + * velocity weighting. This feature is of extreme importance for totally + * removing ghosting. More information here: + * http://iryoku.com/aacourse/downloads/13-Anti-Aliasing-Methods-in-CryENGINE-3.pdf + * + * Note that you'll need to setup a velocity buffer for enabling reprojection. + * For static geometry, saving the previous depth buffer is a viable + * alternative. + */ +#ifndef SMAA_REPROJECTION +#define SMAA_REPROJECTION 0 +#endif + +/** + * SMAA_REPROJECTION_WEIGHT_SCALE controls the velocity weighting. It allows to + * remove ghosting trails behind the moving object, which are not removed by + * just using reprojection. Using low values will exhibit ghosting, while using + * high values will disable temporal supersampling under motion. + * + * Behind the scenes, velocity weighting removes temporal supersampling when + * the velocity of the subsamples differs (meaning they are different objects). + * + * Range: [0, 80] + */ +#ifndef SMAA_REPROJECTION_WEIGHT_SCALE +#define SMAA_REPROJECTION_WEIGHT_SCALE 30.0 +#endif + +/** + * On some compilers, discard cannot be used in vertex shaders. Thus, they need + * to be compiled separately. + */ +#ifndef SMAA_INCLUDE_VS +#define SMAA_INCLUDE_VS 1 +#endif +#ifndef SMAA_INCLUDE_PS +#define SMAA_INCLUDE_PS 1 +#endif + +//----------------------------------------------------------------------------- +// Texture Access Defines + +#ifndef SMAA_AREATEX_SELECT +#if defined(SMAA_HLSL_3) +#define SMAA_AREATEX_SELECT(sample) sample.ra +#else +#define SMAA_AREATEX_SELECT(sample) sample.rg +#endif +#endif + +#ifndef SMAA_SEARCHTEX_SELECT +#define SMAA_SEARCHTEX_SELECT(sample) sample.r +#endif + +#ifndef SMAA_DECODE_VELOCITY +#define SMAA_DECODE_VELOCITY(sample) sample.rg +#endif + +//----------------------------------------------------------------------------- +// Non-Configurable Defines + +#define SMAA_AREATEX_MAX_DISTANCE 16 +#define SMAA_AREATEX_MAX_DISTANCE_DIAG 20 +#define SMAA_AREATEX_PIXEL_SIZE (1.0 / float2(160.0, 560.0)) +#define SMAA_AREATEX_SUBTEX_SIZE (1.0 / 7.0) +#define SMAA_SEARCHTEX_SIZE float2(66.0, 33.0) +#define SMAA_SEARCHTEX_PACKED_SIZE float2(64.0, 16.0) +#define SMAA_CORNER_ROUNDING_NORM (float(SMAA_CORNER_ROUNDING) / 100.0) + +//----------------------------------------------------------------------------- +// Porting Functions + +#if defined(SMAA_HLSL_3) +#define SMAATexture2D(tex) sampler2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) tex2Dlod(tex, float4(coord, 0.0, 0.0)) +#define SMAASampleLevelZeroPoint(tex, coord) tex2Dlod(tex, float4(coord, 0.0, 0.0)) +#define SMAASampleLevelZeroOffset(tex, coord, offset) tex2Dlod(tex, float4(coord + offset * SMAA_RT_METRICS.xy, 0.0, 0.0)) +#define SMAASample(tex, coord) tex2D(tex, coord) +#define SMAASamplePoint(tex, coord) tex2D(tex, coord) +#define SMAASampleOffset(tex, coord, offset) tex2D(tex, coord + offset * SMAA_RT_METRICS.xy) +#define SMAA_FLATTEN [flatten] +#define SMAA_BRANCH [branch] +#endif +#if defined(SMAA_HLSL_4) || defined(SMAA_HLSL_4_1) +SamplerState LinearSampler { Filter = MIN_MAG_LINEAR_MIP_POINT; AddressU = Clamp; AddressV = Clamp; }; +SamplerState PointSampler { Filter = MIN_MAG_MIP_POINT; AddressU = Clamp; AddressV = Clamp; }; +#define SMAATexture2D(tex) Texture2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) tex.SampleLevel(LinearSampler, coord, 0) +#define SMAASampleLevelZeroPoint(tex, coord) tex.SampleLevel(PointSampler, coord, 0) +#define SMAASampleLevelZeroOffset(tex, coord, offset) tex.SampleLevel(LinearSampler, coord, 0, offset) +#define SMAASample(tex, coord) tex.Sample(LinearSampler, coord) +#define SMAASamplePoint(tex, coord) tex.Sample(PointSampler, coord) +#define SMAASampleOffset(tex, coord, offset) tex.Sample(LinearSampler, coord, offset) +#define SMAA_FLATTEN [flatten] +#define SMAA_BRANCH [branch] +#define SMAATexture2DMS2(tex) Texture2DMS tex +#define SMAALoad(tex, pos, sample) tex.Load(pos, sample) +#if defined(SMAA_HLSL_4_1) +#define SMAAGather(tex, coord) tex.Gather(LinearSampler, coord, 0) +#endif +#endif +#if defined(SMAA_GLSL_3) || defined(SMAA_GLSL_4) +#define SMAATexture2D(tex) sampler2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) textureLod(tex, coord, 0.0) +#define SMAASampleLevelZeroPoint(tex, coord) textureLod(tex, coord, 0.0) +#define SMAASampleLevelZeroOffset(tex, coord, offset) textureLodOffset(tex, coord, 0.0, offset) +#define SMAASample(tex, coord) texture(tex, coord) +#define SMAASamplePoint(tex, coord) texture(tex, coord) +#define SMAASampleOffset(tex, coord, offset) texture(tex, coord, offset) +#define SMAA_FLATTEN +#define SMAA_BRANCH +#define lerp(a, b, t) mix(a, b, t) +#define saturate(a) clamp(a, 0.0, 1.0) +#if defined(SMAA_GLSL_4) +#define mad(a, b, c) fma(a, b, c) +#define SMAAGather(tex, coord) textureGather(tex, coord) +#else +#define mad(a, b, c) (a * b + c) +#endif +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define int2 ivec2 +#define int3 ivec3 +#define int4 ivec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#endif + +#if !defined(SMAA_HLSL_3) && !defined(SMAA_HLSL_4) && !defined(SMAA_HLSL_4_1) && !defined(SMAA_GLSL_3) && !defined(SMAA_GLSL_4) && !defined(SMAA_CUSTOM_SL) +#error you must define the shading language: SMAA_HLSL_*, SMAA_GLSL_* or SMAA_CUSTOM_SL +#endif + +//----------------------------------------------------------------------------- +// Misc functions + +/** + * Gathers current pixel, and the top-left neighbors. + */ +float3 SMAAGatherNeighbours(float2 texcoord, + float4 offset[3], + SMAATexture2D(tex)) { + #ifdef SMAAGather + return SMAAGather(tex, texcoord + SMAA_RT_METRICS.xy * float2(-0.5, -0.5)).grb; + #else + float P = SMAASamplePoint(tex, texcoord).r; + float Pleft = SMAASamplePoint(tex, offset[0].xy).r; + float Ptop = SMAASamplePoint(tex, offset[0].zw).r; + return float3(P, Pleft, Ptop); + #endif +} + +/** + * Adjusts the threshold by means of predication. + */ +float2 SMAACalculatePredicatedThreshold(float2 texcoord, + float4 offset[3], + SMAATexture2D(predicationTex)) { + float3 neighbours = SMAAGatherNeighbours(texcoord, offset, SMAATexturePass2D(predicationTex)); + float2 delta = abs(neighbours.xx - neighbours.yz); + float2 edges = step(SMAA_PREDICATION_THRESHOLD, delta); + return SMAA_PREDICATION_SCALE * SMAA_THRESHOLD * (1.0 - SMAA_PREDICATION_STRENGTH * edges); +} + +/** + * Conditional move: + */ +void SMAAMovc(bool2 cond, inout float2 variable, float2 value) { + SMAA_FLATTEN if (cond.x) variable.x = value.x; + SMAA_FLATTEN if (cond.y) variable.y = value.y; +} + +void SMAAMovc(bool4 cond, inout float4 variable, float4 value) { + SMAAMovc(cond.xy, variable.xy, value.xy); + SMAAMovc(cond.zw, variable.zw, value.zw); +} + + +#if SMAA_INCLUDE_VS +//----------------------------------------------------------------------------- +// Vertex Shaders + +/** + * Edge Detection Vertex Shader + */ +void SMAAEdgeDetectionVS(float2 texcoord, + out float4 offset[3]) { + offset[0] = mad(SMAA_RT_METRICS.xyxy, float4(-1.0, 0.0, 0.0, -1.0), texcoord.xyxy); + offset[1] = mad(SMAA_RT_METRICS.xyxy, float4( 1.0, 0.0, 0.0, 1.0), texcoord.xyxy); + offset[2] = mad(SMAA_RT_METRICS.xyxy, float4(-2.0, 0.0, 0.0, -2.0), texcoord.xyxy); +} + +/** + * Blend Weight Calculation Vertex Shader + */ +void SMAABlendingWeightCalculationVS(float2 texcoord, + out float2 pixcoord, + out float4 offset[3]) { + pixcoord = texcoord * SMAA_RT_METRICS.zw; + + // We will use these offsets for the searches later on (see @PSEUDO_GATHER4): + offset[0] = mad(SMAA_RT_METRICS.xyxy, float4(-0.25, -0.125, 1.25, -0.125), texcoord.xyxy); + offset[1] = mad(SMAA_RT_METRICS.xyxy, float4(-0.125, -0.25, -0.125, 1.25), texcoord.xyxy); + + // And these for the searches, they indicate the ends of the loops: + offset[2] = mad(SMAA_RT_METRICS.xxyy, + float4(-2.0, 2.0, -2.0, 2.0) * float(SMAA_MAX_SEARCH_STEPS), + float4(offset[0].xz, offset[1].yw)); +} + +/** + * Neighborhood Blending Vertex Shader + */ +void SMAANeighborhoodBlendingVS(float2 texcoord, + out float4 offset) { + offset = mad(SMAA_RT_METRICS.xyxy, float4( 1.0, 0.0, 0.0, 1.0), texcoord.xyxy); +} +#endif // SMAA_INCLUDE_VS + +#if SMAA_INCLUDE_PS +//----------------------------------------------------------------------------- +// Edge Detection Pixel Shaders (First Pass) + +/** + * Luma Edge Detection + * + * IMPORTANT NOTICE: luma edge detection requires gamma-corrected colors, and + * thus 'colorTex' should be a non-sRGB texture. + */ +float2 SMAALumaEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(colorTex) + #if SMAA_PREDICATION + , SMAATexture2D(predicationTex) + #endif + ) { + // Calculate the threshold: + #if SMAA_PREDICATION + float2 threshold = SMAACalculatePredicatedThreshold(texcoord, offset, SMAATexturePass2D(predicationTex)); + #else + float2 threshold = float2(SMAA_THRESHOLD, SMAA_THRESHOLD); + #endif + + // Calculate lumas: + float3 weights = float3(0.2126, 0.7152, 0.0722); + float L = dot(SMAASamplePoint(colorTex, texcoord).rgb, weights); + + float Lleft = dot(SMAASamplePoint(colorTex, offset[0].xy).rgb, weights); + float Ltop = dot(SMAASamplePoint(colorTex, offset[0].zw).rgb, weights); + + // We do the usual threshold: + float4 delta; + delta.xy = abs(L - float2(Lleft, Ltop)); + float2 edges = step(threshold, delta.xy); + + // Then discard if there is no edge: + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + // Calculate right and bottom deltas: + float Lright = dot(SMAASamplePoint(colorTex, offset[1].xy).rgb, weights); + float Lbottom = dot(SMAASamplePoint(colorTex, offset[1].zw).rgb, weights); + delta.zw = abs(L - float2(Lright, Lbottom)); + + // Calculate the maximum delta in the direct neighborhood: + float2 maxDelta = max(delta.xy, delta.zw); + + // Calculate left-left and top-top deltas: + float Lleftleft = dot(SMAASamplePoint(colorTex, offset[2].xy).rgb, weights); + float Ltoptop = dot(SMAASamplePoint(colorTex, offset[2].zw).rgb, weights); + delta.zw = abs(float2(Lleft, Ltop) - float2(Lleftleft, Ltoptop)); + + // Calculate the final maximum delta: + maxDelta = max(maxDelta.xy, delta.zw); + float finalDelta = max(maxDelta.x, maxDelta.y); + + // Local contrast adaptation: + edges.xy *= step(finalDelta, SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR * delta.xy); + + return edges; +} + +/** + * Color Edge Detection + * + * IMPORTANT NOTICE: color edge detection requires gamma-corrected colors, and + * thus 'colorTex' should be a non-sRGB texture. + */ +float2 SMAAColorEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(colorTex) + #if SMAA_PREDICATION + , SMAATexture2D(predicationTex) + #endif + ) { + // Calculate the threshold: + #if SMAA_PREDICATION + float2 threshold = SMAACalculatePredicatedThreshold(texcoord, offset, predicationTex); + #else + float2 threshold = float2(SMAA_THRESHOLD, SMAA_THRESHOLD); + #endif + + // Calculate color deltas: + float4 delta; + float3 C = SMAASamplePoint(colorTex, texcoord).rgb; + + float3 Cleft = SMAASamplePoint(colorTex, offset[0].xy).rgb; + float3 t = abs(C - Cleft); + delta.x = max(max(t.r, t.g), t.b); + + float3 Ctop = SMAASamplePoint(colorTex, offset[0].zw).rgb; + t = abs(C - Ctop); + delta.y = max(max(t.r, t.g), t.b); + + // We do the usual threshold: + float2 edges = step(threshold, delta.xy); + + // Then discard if there is no edge: + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + // Calculate right and bottom deltas: + float3 Cright = SMAASamplePoint(colorTex, offset[1].xy).rgb; + t = abs(C - Cright); + delta.z = max(max(t.r, t.g), t.b); + + float3 Cbottom = SMAASamplePoint(colorTex, offset[1].zw).rgb; + t = abs(C - Cbottom); + delta.w = max(max(t.r, t.g), t.b); + + // Calculate the maximum delta in the direct neighborhood: + float2 maxDelta = max(delta.xy, delta.zw); + + // Calculate left-left and top-top deltas: + float3 Cleftleft = SMAASamplePoint(colorTex, offset[2].xy).rgb; + t = abs(C - Cleftleft); + delta.z = max(max(t.r, t.g), t.b); + + float3 Ctoptop = SMAASamplePoint(colorTex, offset[2].zw).rgb; + t = abs(C - Ctoptop); + delta.w = max(max(t.r, t.g), t.b); + + // Calculate the final maximum delta: + maxDelta = max(maxDelta.xy, delta.zw); + float finalDelta = max(maxDelta.x, maxDelta.y); + + // Local contrast adaptation: + edges.xy *= step(finalDelta, SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR * delta.xy); + + return edges; +} + +/** + * Depth Edge Detection + */ +float2 SMAADepthEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(depthTex)) { + float3 neighbours = SMAAGatherNeighbours(texcoord, offset, SMAATexturePass2D(depthTex)); + float2 delta = abs(neighbours.xx - float2(neighbours.y, neighbours.z)); + float2 edges = step(SMAA_DEPTH_THRESHOLD, delta); + + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + return edges; +} + +//----------------------------------------------------------------------------- +// Diagonal Search Functions + +#if !defined(SMAA_DISABLE_DIAG_DETECTION) + +/** + * Allows to decode two binary values from a bilinear-filtered access. + */ +float2 SMAADecodeDiagBilinearAccess(float2 e) { + // Bilinear access for fetching 'e' have a 0.25 offset, and we are + // interested in the R and G edges: + // + // +---G---+-------+ + // | x o R x | + // +-------+-------+ + // + // Then, if one of these edge is enabled: + // Red: (0.75 * X + 0.25 * 1) => 0.25 or 1.0 + // Green: (0.75 * 1 + 0.25 * X) => 0.75 or 1.0 + // + // This function will unpack the values (mad + mul + round): + // wolframalpha.com: round(x * abs(5 * x - 5 * 0.75)) plot 0 to 1 + e.r = e.r * abs(5.0 * e.r - 5.0 * 0.75); + return round(e); +} + +float4 SMAADecodeDiagBilinearAccess(float4 e) { + e.rb = e.rb * abs(5.0 * e.rb - 5.0 * 0.75); + return round(e); +} + +/** + * These functions allows to perform diagonal pattern searches. + */ +float2 SMAASearchDiag1(SMAATexture2D(edgesTex), float2 texcoord, float2 dir, out float2 e) { + float4 coord = float4(texcoord, -1.0, 1.0); + float3 t = float3(SMAA_RT_METRICS.xy, 1.0); + while (coord.z < float(SMAA_MAX_SEARCH_STEPS_DIAG - 1) && + coord.w > 0.9) { + coord.xyz = mad(t, float3(dir, 1.0), coord.xyz); + e = SMAASampleLevelZero(edgesTex, coord.xy).rg; + coord.w = dot(e, float2(0.5, 0.5)); + } + return coord.zw; +} + +float2 SMAASearchDiag2(SMAATexture2D(edgesTex), float2 texcoord, float2 dir, out float2 e) { + float4 coord = float4(texcoord, -1.0, 1.0); + coord.x += 0.25 * SMAA_RT_METRICS.x; // See @SearchDiag2Optimization + float3 t = float3(SMAA_RT_METRICS.xy, 1.0); + while (coord.z < float(SMAA_MAX_SEARCH_STEPS_DIAG - 1) && + coord.w > 0.9) { + coord.xyz = mad(t, float3(dir, 1.0), coord.xyz); + + // @SearchDiag2Optimization + // Fetch both edges at once using bilinear filtering: + e = SMAASampleLevelZero(edgesTex, coord.xy).rg; + e = SMAADecodeDiagBilinearAccess(e); + + // Non-optimized version: + // e.g = SMAASampleLevelZero(edgesTex, coord.xy).g; + // e.r = SMAASampleLevelZeroOffset(edgesTex, coord.xy, int2(1, 0)).r; + + coord.w = dot(e, float2(0.5, 0.5)); + } + return coord.zw; +} + +/** + * Similar to SMAAArea, this calculates the area corresponding to a certain + * diagonal distance and crossing edges 'e'. + */ +float2 SMAAAreaDiag(SMAATexture2D(areaTex), float2 dist, float2 e, float offset) { + float2 texcoord = mad(float2(SMAA_AREATEX_MAX_DISTANCE_DIAG, SMAA_AREATEX_MAX_DISTANCE_DIAG), e, dist); + + // We do a scale and bias for mapping to texel space: + texcoord = mad(SMAA_AREATEX_PIXEL_SIZE, texcoord, 0.5 * SMAA_AREATEX_PIXEL_SIZE); + + // Diagonal areas are on the second half of the texture: + texcoord.x += 0.5; + + // Move to proper place, according to the subpixel offset: + texcoord.y += SMAA_AREATEX_SUBTEX_SIZE * offset; + + // Do it! + return SMAA_AREATEX_SELECT(SMAASampleLevelZero(areaTex, texcoord)); +} + +/** + * This searches for diagonal patterns and returns the corresponding weights. + */ +float2 SMAACalculateDiagWeights(SMAATexture2D(edgesTex), SMAATexture2D(areaTex), float2 texcoord, float2 e, float4 subsampleIndices) { + float2 weights = float2(0.0, 0.0); + + // Search for the line ends: + float4 d; + float2 end; + if (e.r > 0.0) { + d.xz = SMAASearchDiag1(SMAATexturePass2D(edgesTex), texcoord, float2(-1.0, 1.0), end); + d.x += float(end.y > 0.9); + } else + d.xz = float2(0.0, 0.0); + d.yw = SMAASearchDiag1(SMAATexturePass2D(edgesTex), texcoord, float2(1.0, -1.0), end); + + SMAA_BRANCH + if (d.x + d.y > 2.0) { // d.x + d.y + 1 > 3 + // Fetch the crossing edges: + float4 coords = mad(float4(-d.x + 0.25, d.x, d.y, -d.y - 0.25), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + float4 c; + c.xy = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).rg; + c.zw = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).rg; + c.yxwz = SMAADecodeDiagBilinearAccess(c.xyzw); + + // Non-optimized version: + // float4 coords = mad(float4(-d.x, d.x, d.y, -d.y), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + // float4 c; + // c.x = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).g; + // c.y = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2( 0, 0)).r; + // c.z = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).g; + // c.w = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, -1)).r; + + // Merge crossing edges at each side into a single value: + float2 cc = mad(float2(2.0, 2.0), c.xz, c.yw); + + // Remove the crossing edge if we didn't found the end of the line: + SMAAMovc(bool2(step(0.9, d.zw)), cc, float2(0.0, 0.0)); + + // Fetch the areas for this line: + weights += SMAAAreaDiag(SMAATexturePass2D(areaTex), d.xy, cc, subsampleIndices.z); + } + + // Search for the line ends: + d.xz = SMAASearchDiag2(SMAATexturePass2D(edgesTex), texcoord, float2(-1.0, -1.0), end); + if (SMAASampleLevelZeroOffset(edgesTex, texcoord, int2(1, 0)).r > 0.0) { + d.yw = SMAASearchDiag2(SMAATexturePass2D(edgesTex), texcoord, float2(1.0, 1.0), end); + d.y += float(end.y > 0.9); + } else + d.yw = float2(0.0, 0.0); + + SMAA_BRANCH + if (d.x + d.y > 2.0) { // d.x + d.y + 1 > 3 + // Fetch the crossing edges: + float4 coords = mad(float4(-d.x, -d.x, d.y, d.y), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + float4 c; + c.x = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).g; + c.y = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2( 0, -1)).r; + c.zw = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).gr; + float2 cc = mad(float2(2.0, 2.0), c.xz, c.yw); + + // Remove the crossing edge if we didn't found the end of the line: + SMAAMovc(bool2(step(0.9, d.zw)), cc, float2(0.0, 0.0)); + + // Fetch the areas for this line: + weights += SMAAAreaDiag(SMAATexturePass2D(areaTex), d.xy, cc, subsampleIndices.w).gr; + } + + return weights; +} +#endif + +//----------------------------------------------------------------------------- +// Horizontal/Vertical Search Functions + +/** + * This allows to determine how much length should we add in the last step + * of the searches. It takes the bilinearly interpolated edge (see + * @PSEUDO_GATHER4), and adds 0, 1 or 2, depending on which edges and + * crossing edges are active. + */ +float SMAASearchLength(SMAATexture2D(searchTex), float2 e, float offset) { + // The texture is flipped vertically, with left and right cases taking half + // of the space horizontally: + float2 scale = SMAA_SEARCHTEX_SIZE * float2(0.5, -1.0); + float2 bias = SMAA_SEARCHTEX_SIZE * float2(offset, 1.0); + + // Scale and bias to access texel centers: + scale += float2(-1.0, 1.0); + bias += float2( 0.5, -0.5); + + // Convert from pixel coordinates to texcoords: + // (We use SMAA_SEARCHTEX_PACKED_SIZE because the texture is cropped) + scale *= 1.0 / SMAA_SEARCHTEX_PACKED_SIZE; + bias *= 1.0 / SMAA_SEARCHTEX_PACKED_SIZE; + + // Lookup the search texture: + return SMAA_SEARCHTEX_SELECT(SMAASampleLevelZero(searchTex, mad(scale, e, bias))); +} + +/** + * Horizontal/vertical search functions for the 2nd pass. + */ +float SMAASearchXLeft(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + /** + * @PSEUDO_GATHER4 + * This texcoord has been offset by (-0.25, -0.125) in the vertex shader to + * sample between edge, thus fetching four edges in a row. + * Sampling with different offsets in each direction allows to disambiguate + * which edges are active from the four fetched ones. + */ + float2 e = float2(0.0, 1.0); + while (texcoord.x > end && + e.g > 0.8281 && // Is there some edge not activated? + e.r == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(-float2(2.0, 0.0), SMAA_RT_METRICS.xy, texcoord); + } + + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.0), 3.25); + return mad(SMAA_RT_METRICS.x, offset, texcoord.x); + + // Non-optimized version: + // We correct the previous (-0.25, -0.125) offset we applied: + // texcoord.x += 0.25 * SMAA_RT_METRICS.x; + + // The searches are bias by 1, so adjust the coords accordingly: + // texcoord.x += SMAA_RT_METRICS.x; + + // Disambiguate the length added by the last step: + // texcoord.x += 2.0 * SMAA_RT_METRICS.x; // Undo last step + // texcoord.x -= SMAA_RT_METRICS.x * (255.0 / 127.0) * SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.0); + // return mad(SMAA_RT_METRICS.x, offset, texcoord.x); +} + +float SMAASearchXRight(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(0.0, 1.0); + while (texcoord.x < end && + e.g > 0.8281 && // Is there some edge not activated? + e.r == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(float2(2.0, 0.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.5), 3.25); + return mad(-SMAA_RT_METRICS.x, offset, texcoord.x); +} + +float SMAASearchYUp(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(1.0, 0.0); + while (texcoord.y > end && + e.r > 0.8281 && // Is there some edge not activated? + e.g == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(-float2(0.0, 2.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e.gr, 0.0), 3.25); + return mad(SMAA_RT_METRICS.y, offset, texcoord.y); +} + +float SMAASearchYDown(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(1.0, 0.0); + while (texcoord.y < end && + e.r > 0.8281 && // Is there some edge not activated? + e.g == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(float2(0.0, 2.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e.gr, 0.5), 3.25); + return mad(-SMAA_RT_METRICS.y, offset, texcoord.y); +} + +/** + * Ok, we have the distance and both crossing edges. So, what are the areas + * at each side of current edge? + */ +float2 SMAAArea(SMAATexture2D(areaTex), float2 dist, float e1, float e2, float offset) { + // Rounding prevents precision errors of bilinear filtering: + float2 texcoord = mad(float2(SMAA_AREATEX_MAX_DISTANCE, SMAA_AREATEX_MAX_DISTANCE), round(4.0 * float2(e1, e2)), dist); + + // We do a scale and bias for mapping to texel space: + texcoord = mad(SMAA_AREATEX_PIXEL_SIZE, texcoord, 0.5 * SMAA_AREATEX_PIXEL_SIZE); + + // Move to proper place, according to the subpixel offset: + texcoord.y = mad(SMAA_AREATEX_SUBTEX_SIZE, offset, texcoord.y); + + // Do it! + return SMAA_AREATEX_SELECT(SMAASampleLevelZero(areaTex, texcoord)); +} + +//----------------------------------------------------------------------------- +// Corner Detection Functions + +void SMAADetectHorizontalCornerPattern(SMAATexture2D(edgesTex), inout float2 weights, float4 texcoord, float2 d) { + #if !defined(SMAA_DISABLE_CORNER_DETECTION) + float2 leftRight = step(d.xy, d.yx); + float2 rounding = (1.0 - SMAA_CORNER_ROUNDING_NORM) * leftRight; + + rounding /= leftRight.x + leftRight.y; // Reduce blending for pixels in the center of a line. + + float2 factor = float2(1.0, 1.0); + factor.x -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(0, 1)).r; + factor.x -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(1, 1)).r; + factor.y -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(0, -2)).r; + factor.y -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(1, -2)).r; + + weights *= saturate(factor); + #endif +} + +void SMAADetectVerticalCornerPattern(SMAATexture2D(edgesTex), inout float2 weights, float4 texcoord, float2 d) { + #if !defined(SMAA_DISABLE_CORNER_DETECTION) + float2 leftRight = step(d.xy, d.yx); + float2 rounding = (1.0 - SMAA_CORNER_ROUNDING_NORM) * leftRight; + + rounding /= leftRight.x + leftRight.y; + + float2 factor = float2(1.0, 1.0); + factor.x -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2( 1, 0)).g; + factor.x -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2( 1, 1)).g; + factor.y -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(-2, 0)).g; + factor.y -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(-2, 1)).g; + + weights *= saturate(factor); + #endif +} + +//----------------------------------------------------------------------------- +// Blending Weight Calculation Pixel Shader (Second Pass) + +float4 SMAABlendingWeightCalculationPS(float2 texcoord, + float2 pixcoord, + float4 offset[3], + SMAATexture2D(edgesTex), + SMAATexture2D(areaTex), + SMAATexture2D(searchTex), + float4 subsampleIndices) { // Just pass zero for SMAA 1x, see @SUBSAMPLE_INDICES. + float4 weights = float4(0.0, 0.0, 0.0, 0.0); + + float2 e = SMAASample(edgesTex, texcoord).rg; + + SMAA_BRANCH + if (e.g > 0.0) { // Edge at north + #if !defined(SMAA_DISABLE_DIAG_DETECTION) + // Diagonals have both north and west edges, so searching for them in + // one of the boundaries is enough. + weights.rg = SMAACalculateDiagWeights(SMAATexturePass2D(edgesTex), SMAATexturePass2D(areaTex), texcoord, e, subsampleIndices); + + // We give priority to diagonals, so if we find a diagonal we skip + // horizontal/vertical processing. + SMAA_BRANCH + if (weights.r == -weights.g) { // weights.r + weights.g == 0.0 + #endif + + float2 d; + + // Find the distance to the left: + float3 coords; + coords.x = SMAASearchXLeft(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[0].xy, offset[2].x); + coords.y = offset[1].y; // offset[1].y = texcoord.y - 0.25 * SMAA_RT_METRICS.y (@CROSSING_OFFSET) + d.x = coords.x; + + // Now fetch the left crossing edges, two at a time using bilinear + // filtering. Sampling at -0.25 (see @CROSSING_OFFSET) enables to + // discern what value each edge has: + float e1 = SMAASampleLevelZero(edgesTex, coords.xy).r; + + // Find the distance to the right: + coords.z = SMAASearchXRight(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[0].zw, offset[2].y); + d.y = coords.z; + + // We want the distances to be in pixel units (doing this here allow to + // better interleave arithmetic and memory accesses): + d = abs(round(mad(SMAA_RT_METRICS.zz, d, -pixcoord.xx))); + + // SMAAArea below needs a sqrt, as the areas texture is compressed + // quadratically: + float2 sqrt_d = sqrt(d); + + // Fetch the right crossing edges: + float e2 = SMAASampleLevelZeroOffset(edgesTex, coords.zy, int2(1, 0)).r; + + // Ok, we know how this pattern looks like, now it is time for getting + // the actual area: + weights.rg = SMAAArea(SMAATexturePass2D(areaTex), sqrt_d, e1, e2, subsampleIndices.y); + + // Fix corners: + coords.y = texcoord.y; + SMAADetectHorizontalCornerPattern(SMAATexturePass2D(edgesTex), weights.rg, coords.xyzy, d); + + #if !defined(SMAA_DISABLE_DIAG_DETECTION) + } else + e.r = 0.0; // Skip vertical processing. + #endif + } + + SMAA_BRANCH + if (e.r > 0.0) { // Edge at west + float2 d; + + // Find the distance to the top: + float3 coords; + coords.y = SMAASearchYUp(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[1].xy, offset[2].z); + coords.x = offset[0].x; // offset[1].x = texcoord.x - 0.25 * SMAA_RT_METRICS.x; + d.x = coords.y; + + // Fetch the top crossing edges: + float e1 = SMAASampleLevelZero(edgesTex, coords.xy).g; + + // Find the distance to the bottom: + coords.z = SMAASearchYDown(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[1].zw, offset[2].w); + d.y = coords.z; + + // We want the distances to be in pixel units: + d = abs(round(mad(SMAA_RT_METRICS.ww, d, -pixcoord.yy))); + + // SMAAArea below needs a sqrt, as the areas texture is compressed + // quadratically: + float2 sqrt_d = sqrt(d); + + // Fetch the bottom crossing edges: + float e2 = SMAASampleLevelZeroOffset(edgesTex, coords.xz, int2(0, 1)).g; + + // Get the area for this direction: + weights.ba = SMAAArea(SMAATexturePass2D(areaTex), sqrt_d, e1, e2, subsampleIndices.x); + + // Fix corners: + coords.x = texcoord.x; + SMAADetectVerticalCornerPattern(SMAATexturePass2D(edgesTex), weights.ba, coords.xyxz, d); + } + + return weights; +} + +//----------------------------------------------------------------------------- +// Neighborhood Blending Pixel Shader (Third Pass) + +float4 SMAANeighborhoodBlendingPS(float2 texcoord, + float4 offset, + SMAATexture2D(colorTex), + SMAATexture2D(blendTex) + #if SMAA_REPROJECTION + , SMAATexture2D(velocityTex) + #endif + ) { + // Fetch the blending weights for current pixel: + float4 a; + a.x = SMAASample(blendTex, offset.xy).a; // Right + a.y = SMAASample(blendTex, offset.zw).g; // Top + a.wz = SMAASample(blendTex, texcoord).xz; // Bottom / Left + + // Is there any blending weight with a value greater than 0.0? + SMAA_BRANCH + if (dot(a, float4(1.0, 1.0, 1.0, 1.0)) < 1e-5) { + float4 color = SMAASampleLevelZero(colorTex, texcoord); + + #if SMAA_REPROJECTION + float2 velocity = SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, texcoord)); + + // Pack velocity into the alpha channel: + color.a = sqrt(5.0 * length(velocity)); + #endif + + return color; + } else { + bool h = max(a.x, a.z) > max(a.y, a.w); // max(horizontal) > max(vertical) + + // Calculate the blending offsets: + float4 blendingOffset = float4(0.0, a.y, 0.0, a.w); + float2 blendingWeight = a.yw; + SMAAMovc(bool4(h, h, h, h), blendingOffset, float4(a.x, 0.0, a.z, 0.0)); + SMAAMovc(bool2(h, h), blendingWeight, a.xz); + blendingWeight /= dot(blendingWeight, float2(1.0, 1.0)); + + // Calculate the texture coordinates: + float4 blendingCoord = mad(blendingOffset, float4(SMAA_RT_METRICS.xy, -SMAA_RT_METRICS.xy), texcoord.xyxy); + + // We exploit bilinear filtering to mix current pixel with the chosen + // neighbor: + float4 color = blendingWeight.x * SMAASampleLevelZero(colorTex, blendingCoord.xy); + color += blendingWeight.y * SMAASampleLevelZero(colorTex, blendingCoord.zw); + + #if SMAA_REPROJECTION + // Antialias velocity for proper reprojection in a later stage: + float2 velocity = blendingWeight.x * SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, blendingCoord.xy)); + velocity += blendingWeight.y * SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, blendingCoord.zw)); + + // Pack velocity into the alpha channel: + color.a = sqrt(5.0 * length(velocity)); + #endif + + return color; + } +} + +//----------------------------------------------------------------------------- +// Temporal Resolve Pixel Shader (Optional Pass) + +float4 SMAAResolvePS(float2 texcoord, + SMAATexture2D(currentColorTex), + SMAATexture2D(previousColorTex) + #if SMAA_REPROJECTION + , SMAATexture2D(velocityTex) + #endif + ) { + #if SMAA_REPROJECTION + // Velocity is assumed to be calculated for motion blur, so we need to + // inverse it for reprojection: + float2 velocity = -SMAA_DECODE_VELOCITY(SMAASamplePoint(velocityTex, texcoord).rg); + + // Fetch current pixel: + float4 current = SMAASamplePoint(currentColorTex, texcoord); + + // Reproject current coordinates and fetch previous pixel: + float4 previous = SMAASamplePoint(previousColorTex, texcoord + velocity); + + // Attenuate the previous pixel if the velocity is different: + float delta = abs(current.a * current.a - previous.a * previous.a) / 5.0; + float weight = 0.5 * saturate(1.0 - sqrt(delta) * SMAA_REPROJECTION_WEIGHT_SCALE); + + // Blend the pixels according to the calculated weight: + return lerp(current, previous, weight); + #else + // Just blend the pixels: + float4 current = SMAASamplePoint(currentColorTex, texcoord); + float4 previous = SMAASamplePoint(previousColorTex, texcoord); + return lerp(current, previous, 0.5); + #endif +} + +//----------------------------------------------------------------------------- +// Separate Multisamples Pixel Shader (Optional Pass) + +#ifdef SMAALoad +void SMAASeparatePS(float4 position, + float2 texcoord, + out float4 target0, + out float4 target1, + SMAATexture2DMS2(colorTexMS)) { + int2 pos = int2(position.xy); + target0 = SMAALoad(colorTexMS, pos, 0); + target1 = SMAALoad(colorTexMS, pos, 1); +} +#endif + +//----------------------------------------------------------------------------- +#endif // SMAA_INCLUDE_PS + +layout(rgba8, binding = 0, set = 3) uniform image2D imgOutput; + +layout(binding = 1, set = 2) uniform sampler2D inputImg; +layout( binding = 2 ) uniform invResolution +{ + vec2 invResolution_data; +}; + +void main() +{ + vec2 loc = ivec2(gl_GlobalInvocationID.x * 4, gl_GlobalInvocationID.y * 4); + for(int i = 0; i < 4; i++) + { + for(int j = 0; j < 4; j++) + { + ivec2 texelCoord = ivec2(loc.x + i, loc.y + j); + vec2 coord = (texelCoord + vec2(0.5)) / invResolution_data; + vec4 offset[3]; + SMAAEdgeDetectionVS(coord, offset); + vec2 oColor = SMAAColorEdgeDetectionPS(coord, offset, inputImg); + if (oColor != float2(-2.0, -2.0)) + { + imageStore(imgOutput, texelCoord, vec4(oColor, 0.0, 1.0)); + } + } + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaEdge.spv b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaEdge.spv new file mode 100644 index 00000000..1062a9e3 Binary files /dev/null and b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaEdge.spv differ diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaNeighbour.glsl b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaNeighbour.glsl new file mode 100644 index 00000000..df30d727 --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaNeighbour.glsl @@ -0,0 +1,1403 @@ +#version 430 core +#define SMAA_GLSL_4 1 + +layout (constant_id = 0) const int SMAA_PRESET_LOW = 0; +layout (constant_id = 1) const int SMAA_PRESET_MEDIUM = 0; +layout (constant_id = 2) const int SMAA_PRESET_HIGH = 0; +layout (constant_id = 3) const int SMAA_PRESET_ULTRA = 0; +layout (constant_id = 4) const float METRIC_WIDTH = 1920.0; +layout (constant_id = 5) const float METRIC_HEIGHT = 1080.0; + +#define SMAA_RT_METRICS float4(1.0 / METRIC_WIDTH, 1.0 / METRIC_HEIGHT, METRIC_WIDTH, METRIC_HEIGHT) + +layout (local_size_x = 16, local_size_y = 16) in; +/** + * Copyright (C) 2013 Jorge Jimenez (jorge@iryoku.com) + * Copyright (C) 2013 Jose I. Echevarria (joseignacioechevarria@gmail.com) + * Copyright (C) 2013 Belen Masia (bmasia@unizar.es) + * Copyright (C) 2013 Fernando Navarro (fernandn@microsoft.com) + * Copyright (C) 2013 Diego Gutierrez (diegog@unizar.es) + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * this software and associated documentation files (the "Software"), to deal in + * the Software without restriction, including without limitation the rights to + * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies + * of the Software, and to permit persons to whom the Software is furnished to + * do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. As clarification, there + * is no requirement that the copyright notice and permission be included in + * binary distributions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + + +/** + * _______ ___ ___ ___ ___ + * / || \/ | / \ / \ + * | (---- | \ / | / ^ \ / ^ \ + * \ \ | |\/| | / /_\ \ / /_\ \ + * ----) | | | | | / _____ \ / _____ \ + * |_______/ |__| |__| /__/ \__\ /__/ \__\ + * + * E N H A N C E D + * S U B P I X E L M O R P H O L O G I C A L A N T I A L I A S I N G + * + * http://www.iryoku.com/smaa/ + * + * Hi, welcome aboard! + * + * Here you'll find instructions to get the shader up and running as fast as + * possible. + * + * IMPORTANTE NOTICE: when updating, remember to update both this file and the + * precomputed textures! They may change from version to version. + * + * The shader has three passes, chained together as follows: + * + * |input|------------------ + * v | + * [ SMAA*EdgeDetection ] | + * v | + * |edgesTex| | + * v | + * [ SMAABlendingWeightCalculation ] | + * v | + * |blendTex| | + * v | + * [ SMAANeighborhoodBlending ] <------ + * v + * |output| + * + * Note that each [pass] has its own vertex and pixel shader. Remember to use + * oversized triangles instead of quads to avoid overshading along the + * diagonal. + * + * You've three edge detection methods to choose from: luma, color or depth. + * They represent different quality/performance and anti-aliasing/sharpness + * tradeoffs, so our recommendation is for you to choose the one that best + * suits your particular scenario: + * + * - Depth edge detection is usually the fastest but it may miss some edges. + * + * - Luma edge detection is usually more expensive than depth edge detection, + * but catches visible edges that depth edge detection can miss. + * + * - Color edge detection is usually the most expensive one but catches + * chroma-only edges. + * + * For quickstarters: just use luma edge detection. + * + * The general advice is to not rush the integration process and ensure each + * step is done correctly (don't try to integrate SMAA T2x with predicated edge + * detection from the start!). Ok then, let's go! + * + * 1. The first step is to create two RGBA temporal render targets for holding + * |edgesTex| and |blendTex|. + * + * In DX10 or DX11, you can use a RG render target for the edges texture. + * In the case of NVIDIA GPUs, using RG render targets seems to actually be + * slower. + * + * On the Xbox 360, you can use the same render target for resolving both + * |edgesTex| and |blendTex|, as they aren't needed simultaneously. + * + * 2. Both temporal render targets |edgesTex| and |blendTex| must be cleared + * each frame. Do not forget to clear the alpha channel! + * + * 3. The next step is loading the two supporting precalculated textures, + * 'areaTex' and 'searchTex'. You'll find them in the 'Textures' folder as + * C++ headers, and also as regular DDS files. They'll be needed for the + * 'SMAABlendingWeightCalculation' pass. + * + * If you use the C++ headers, be sure to load them in the format specified + * inside of them. + * + * You can also compress 'areaTex' and 'searchTex' using BC5 and BC4 + * respectively, if you have that option in your content processor pipeline. + * When compressing then, you get a non-perceptible quality decrease, and a + * marginal performance increase. + * + * 4. All samplers must be set to linear filtering and clamp. + * + * After you get the technique working, remember that 64-bit inputs have + * half-rate linear filtering on GCN. + * + * If SMAA is applied to 64-bit color buffers, switching to point filtering + * when accesing them will increase the performance. Search for + * 'SMAASamplePoint' to see which textures may benefit from point + * filtering, and where (which is basically the color input in the edge + * detection and resolve passes). + * + * 5. All texture reads and buffer writes must be non-sRGB, with the exception + * of the input read and the output write in + * 'SMAANeighborhoodBlending' (and only in this pass!). If sRGB reads in + * this last pass are not possible, the technique will work anyway, but + * will perform antialiasing in gamma space. + * + * IMPORTANT: for best results the input read for the color/luma edge + * detection should *NOT* be sRGB. + * + * 6. Before including SMAA.h you'll have to setup the render target metrics, + * the target and any optional configuration defines. Optionally you can + * use a preset. + * + * You have the following targets available: + * SMAA_HLSL_3 + * SMAA_HLSL_4 + * SMAA_HLSL_4_1 + * SMAA_GLSL_3 * + * SMAA_GLSL_4 * + * + * * (See SMAA_INCLUDE_VS and SMAA_INCLUDE_PS below). + * + * And four presets: + * SMAA_PRESET_LOW (%60 of the quality) + * SMAA_PRESET_MEDIUM (%80 of the quality) + * SMAA_PRESET_HIGH (%95 of the quality) + * SMAA_PRESET_ULTRA (%99 of the quality) + * + * For example: + * #define SMAA_RT_METRICS float4(1.0 / 1280.0, 1.0 / 720.0, 1280.0, 720.0) + * #define SMAA_HLSL_4 + * #define SMAA_PRESET_HIGH + * #include "SMAA.h" + * + * Note that SMAA_RT_METRICS doesn't need to be a macro, it can be a + * uniform variable. The code is designed to minimize the impact of not + * using a constant value, but it is still better to hardcode it. + * + * Depending on how you encoded 'areaTex' and 'searchTex', you may have to + * add (and customize) the following defines before including SMAA.h: + * #define SMAA_AREATEX_SELECT(sample) sample.rg + * #define SMAA_SEARCHTEX_SELECT(sample) sample.r + * + * If your engine is already using porting macros, you can define + * SMAA_CUSTOM_SL, and define the porting functions by yourself. + * + * 7. Then, you'll have to setup the passes as indicated in the scheme above. + * You can take a look into SMAA.fx, to see how we did it for our demo. + * Checkout the function wrappers, you may want to copy-paste them! + * + * 8. It's recommended to validate the produced |edgesTex| and |blendTex|. + * You can use a screenshot from your engine to compare the |edgesTex| + * and |blendTex| produced inside of the engine with the results obtained + * with the reference demo. + * + * 9. After you get the last pass to work, it's time to optimize. You'll have + * to initialize a stencil buffer in the first pass (discard is already in + * the code), then mask execution by using it the second pass. The last + * pass should be executed in all pixels. + * + * + * After this point you can choose to enable predicated thresholding, + * temporal supersampling and motion blur integration: + * + * a) If you want to use predicated thresholding, take a look into + * SMAA_PREDICATION; you'll need to pass an extra texture in the edge + * detection pass. + * + * b) If you want to enable temporal supersampling (SMAA T2x): + * + * 1. The first step is to render using subpixel jitters. I won't go into + * detail, but it's as simple as moving each vertex position in the + * vertex shader, you can check how we do it in our DX10 demo. + * + * 2. Then, you must setup the temporal resolve. You may want to take a look + * into SMAAResolve for resolving 2x modes. After you get it working, you'll + * probably see ghosting everywhere. But fear not, you can enable the + * CryENGINE temporal reprojection by setting the SMAA_REPROJECTION macro. + * Check out SMAA_DECODE_VELOCITY if your velocity buffer is encoded. + * + * 3. The next step is to apply SMAA to each subpixel jittered frame, just as + * done for 1x. + * + * 4. At this point you should already have something usable, but for best + * results the proper area textures must be set depending on current jitter. + * For this, the parameter 'subsampleIndices' of + * 'SMAABlendingWeightCalculationPS' must be set as follows, for our T2x + * mode: + * + * @SUBSAMPLE_INDICES + * + * | S# | Camera Jitter | subsampleIndices | + * +----+------------------+---------------------+ + * | 0 | ( 0.25, -0.25) | float4(1, 1, 1, 0) | + * | 1 | (-0.25, 0.25) | float4(2, 2, 2, 0) | + * + * These jitter positions assume a bottom-to-top y axis. S# stands for the + * sample number. + * + * More information about temporal supersampling here: + * http://iryoku.com/aacourse/downloads/13-Anti-Aliasing-Methods-in-CryENGINE-3.pdf + * + * c) If you want to enable spatial multisampling (SMAA S2x): + * + * 1. The scene must be rendered using MSAA 2x. The MSAA 2x buffer must be + * created with: + * - DX10: see below (*) + * - DX10.1: D3D10_STANDARD_MULTISAMPLE_PATTERN or + * - DX11: D3D11_STANDARD_MULTISAMPLE_PATTERN + * + * This allows to ensure that the subsample order matches the table in + * @SUBSAMPLE_INDICES. + * + * (*) In the case of DX10, we refer the reader to: + * - SMAA::detectMSAAOrder and + * - SMAA::msaaReorder + * + * These functions allow to match the standard multisample patterns by + * detecting the subsample order for a specific GPU, and reordering + * them appropriately. + * + * 2. A shader must be run to output each subsample into a separate buffer + * (DX10 is required). You can use SMAASeparate for this purpose, or just do + * it in an existing pass (for example, in the tone mapping pass, which has + * the advantage of feeding tone mapped subsamples to SMAA, which will yield + * better results). + * + * 3. The full SMAA 1x pipeline must be run for each separated buffer, storing + * the results in the final buffer. The second run should alpha blend with + * the existing final buffer using a blending factor of 0.5. + * 'subsampleIndices' must be adjusted as in the SMAA T2x case (see point + * b). + * + * d) If you want to enable temporal supersampling on top of SMAA S2x + * (which actually is SMAA 4x): + * + * 1. SMAA 4x consists on temporally jittering SMAA S2x, so the first step is + * to calculate SMAA S2x for current frame. In this case, 'subsampleIndices' + * must be set as follows: + * + * | F# | S# | Camera Jitter | Net Jitter | subsampleIndices | + * +----+----+--------------------+-------------------+----------------------+ + * | 0 | 0 | ( 0.125, 0.125) | ( 0.375, -0.125) | float4(5, 3, 1, 3) | + * | 0 | 1 | ( 0.125, 0.125) | (-0.125, 0.375) | float4(4, 6, 2, 3) | + * +----+----+--------------------+-------------------+----------------------+ + * | 1 | 2 | (-0.125, -0.125) | ( 0.125, -0.375) | float4(3, 5, 1, 4) | + * | 1 | 3 | (-0.125, -0.125) | (-0.375, 0.125) | float4(6, 4, 2, 4) | + * + * These jitter positions assume a bottom-to-top y axis. F# stands for the + * frame number. S# stands for the sample number. + * + * 2. After calculating SMAA S2x for current frame (with the new subsample + * indices), previous frame must be reprojected as in SMAA T2x mode (see + * point b). + * + * e) If motion blur is used, you may want to do the edge detection pass + * together with motion blur. This has two advantages: + * + * 1. Pixels under heavy motion can be omitted from the edge detection process. + * For these pixels we can just store "no edge", as motion blur will take + * care of them. + * 2. The center pixel tap is reused. + * + * Note that in this case depth testing should be used instead of stenciling, + * as we have to write all the pixels in the motion blur pass. + * + * That's it! + */ + +//----------------------------------------------------------------------------- +// SMAA Presets + +/** + * Note that if you use one of these presets, the following configuration + * macros will be ignored if set in the "Configurable Defines" section. + */ + +#if defined(SMAA_PRESET_LOW) +#define SMAA_THRESHOLD 0.15 +#define SMAA_MAX_SEARCH_STEPS 4 +#define SMAA_DISABLE_DIAG_DETECTION +#define SMAA_DISABLE_CORNER_DETECTION +#elif defined(SMAA_PRESET_MEDIUM) +#define SMAA_THRESHOLD 0.1 +#define SMAA_MAX_SEARCH_STEPS 8 +#define SMAA_DISABLE_DIAG_DETECTION +#define SMAA_DISABLE_CORNER_DETECTION +#elif defined(SMAA_PRESET_HIGH) +#define SMAA_THRESHOLD 0.1 +#define SMAA_MAX_SEARCH_STEPS 16 +#define SMAA_MAX_SEARCH_STEPS_DIAG 8 +#define SMAA_CORNER_ROUNDING 25 +#elif defined(SMAA_PRESET_ULTRA) +#define SMAA_THRESHOLD 0.05 +#define SMAA_MAX_SEARCH_STEPS 32 +#define SMAA_MAX_SEARCH_STEPS_DIAG 16 +#define SMAA_CORNER_ROUNDING 25 +#endif + +//----------------------------------------------------------------------------- +// Configurable Defines + +/** + * SMAA_THRESHOLD specifies the threshold or sensitivity to edges. + * Lowering this value you will be able to detect more edges at the expense of + * performance. + * + * Range: [0, 0.5] + * 0.1 is a reasonable value, and allows to catch most visible edges. + * 0.05 is a rather overkill value, that allows to catch 'em all. + * + * If temporal supersampling is used, 0.2 could be a reasonable value, as low + * contrast edges are properly filtered by just 2x. + */ +#ifndef SMAA_THRESHOLD +#define SMAA_THRESHOLD 0.1 +#endif + +/** + * SMAA_DEPTH_THRESHOLD specifies the threshold for depth edge detection. + * + * Range: depends on the depth range of the scene. + */ +#ifndef SMAA_DEPTH_THRESHOLD +#define SMAA_DEPTH_THRESHOLD (0.1 * SMAA_THRESHOLD) +#endif + +/** + * SMAA_MAX_SEARCH_STEPS specifies the maximum steps performed in the + * horizontal/vertical pattern searches, at each side of the pixel. + * + * In number of pixels, it's actually the double. So the maximum line length + * perfectly handled by, for example 16, is 64 (by perfectly, we meant that + * longer lines won't look as good, but still antialiased). + * + * Range: [0, 112] + */ +#ifndef SMAA_MAX_SEARCH_STEPS +#define SMAA_MAX_SEARCH_STEPS 16 +#endif + +/** + * SMAA_MAX_SEARCH_STEPS_DIAG specifies the maximum steps performed in the + * diagonal pattern searches, at each side of the pixel. In this case we jump + * one pixel at time, instead of two. + * + * Range: [0, 20] + * + * On high-end machines it is cheap (between a 0.8x and 0.9x slower for 16 + * steps), but it can have a significant impact on older machines. + * + * Define SMAA_DISABLE_DIAG_DETECTION to disable diagonal processing. + */ +#ifndef SMAA_MAX_SEARCH_STEPS_DIAG +#define SMAA_MAX_SEARCH_STEPS_DIAG 8 +#endif + +/** + * SMAA_CORNER_ROUNDING specifies how much sharp corners will be rounded. + * + * Range: [0, 100] + * + * Define SMAA_DISABLE_CORNER_DETECTION to disable corner processing. + */ +#ifndef SMAA_CORNER_ROUNDING +#define SMAA_CORNER_ROUNDING 25 +#endif + +/** + * If there is an neighbor edge that has SMAA_LOCAL_CONTRAST_FACTOR times + * bigger contrast than current edge, current edge will be discarded. + * + * This allows to eliminate spurious crossing edges, and is based on the fact + * that, if there is too much contrast in a direction, that will hide + * perceptually contrast in the other neighbors. + */ +#ifndef SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR +#define SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR 2.0 +#endif + +/** + * Predicated thresholding allows to better preserve texture details and to + * improve performance, by decreasing the number of detected edges using an + * additional buffer like the light accumulation buffer, object ids or even the + * depth buffer (the depth buffer usage may be limited to indoor or short range + * scenes). + * + * It locally decreases the luma or color threshold if an edge is found in an + * additional buffer (so the global threshold can be higher). + * + * This method was developed by Playstation EDGE MLAA team, and used in + * Killzone 3, by using the light accumulation buffer. More information here: + * http://iryoku.com/aacourse/downloads/06-MLAA-on-PS3.pptx + */ +#ifndef SMAA_PREDICATION +#define SMAA_PREDICATION 0 +#endif + +/** + * Threshold to be used in the additional predication buffer. + * + * Range: depends on the input, so you'll have to find the magic number that + * works for you. + */ +#ifndef SMAA_PREDICATION_THRESHOLD +#define SMAA_PREDICATION_THRESHOLD 0.01 +#endif + +/** + * How much to scale the global threshold used for luma or color edge + * detection when using predication. + * + * Range: [1, 5] + */ +#ifndef SMAA_PREDICATION_SCALE +#define SMAA_PREDICATION_SCALE 2.0 +#endif + +/** + * How much to locally decrease the threshold. + * + * Range: [0, 1] + */ +#ifndef SMAA_PREDICATION_STRENGTH +#define SMAA_PREDICATION_STRENGTH 0.4 +#endif + +/** + * Temporal reprojection allows to remove ghosting artifacts when using + * temporal supersampling. We use the CryEngine 3 method which also introduces + * velocity weighting. This feature is of extreme importance for totally + * removing ghosting. More information here: + * http://iryoku.com/aacourse/downloads/13-Anti-Aliasing-Methods-in-CryENGINE-3.pdf + * + * Note that you'll need to setup a velocity buffer for enabling reprojection. + * For static geometry, saving the previous depth buffer is a viable + * alternative. + */ +#ifndef SMAA_REPROJECTION +#define SMAA_REPROJECTION 0 +#endif + +/** + * SMAA_REPROJECTION_WEIGHT_SCALE controls the velocity weighting. It allows to + * remove ghosting trails behind the moving object, which are not removed by + * just using reprojection. Using low values will exhibit ghosting, while using + * high values will disable temporal supersampling under motion. + * + * Behind the scenes, velocity weighting removes temporal supersampling when + * the velocity of the subsamples differs (meaning they are different objects). + * + * Range: [0, 80] + */ +#ifndef SMAA_REPROJECTION_WEIGHT_SCALE +#define SMAA_REPROJECTION_WEIGHT_SCALE 30.0 +#endif + +/** + * On some compilers, discard cannot be used in vertex shaders. Thus, they need + * to be compiled separately. + */ +#ifndef SMAA_INCLUDE_VS +#define SMAA_INCLUDE_VS 1 +#endif +#ifndef SMAA_INCLUDE_PS +#define SMAA_INCLUDE_PS 1 +#endif + +//----------------------------------------------------------------------------- +// Texture Access Defines + +#ifndef SMAA_AREATEX_SELECT +#if defined(SMAA_HLSL_3) +#define SMAA_AREATEX_SELECT(sample) sample.ra +#else +#define SMAA_AREATEX_SELECT(sample) sample.rg +#endif +#endif + +#ifndef SMAA_SEARCHTEX_SELECT +#define SMAA_SEARCHTEX_SELECT(sample) sample.r +#endif + +#ifndef SMAA_DECODE_VELOCITY +#define SMAA_DECODE_VELOCITY(sample) sample.rg +#endif + +//----------------------------------------------------------------------------- +// Non-Configurable Defines + +#define SMAA_AREATEX_MAX_DISTANCE 16 +#define SMAA_AREATEX_MAX_DISTANCE_DIAG 20 +#define SMAA_AREATEX_PIXEL_SIZE (1.0 / float2(160.0, 560.0)) +#define SMAA_AREATEX_SUBTEX_SIZE (1.0 / 7.0) +#define SMAA_SEARCHTEX_SIZE float2(66.0, 33.0) +#define SMAA_SEARCHTEX_PACKED_SIZE float2(64.0, 16.0) +#define SMAA_CORNER_ROUNDING_NORM (float(SMAA_CORNER_ROUNDING) / 100.0) + +//----------------------------------------------------------------------------- +// Porting Functions + +#if defined(SMAA_HLSL_3) +#define SMAATexture2D(tex) sampler2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) tex2Dlod(tex, float4(coord, 0.0, 0.0)) +#define SMAASampleLevelZeroPoint(tex, coord) tex2Dlod(tex, float4(coord, 0.0, 0.0)) +#define SMAASampleLevelZeroOffset(tex, coord, offset) tex2Dlod(tex, float4(coord + offset * SMAA_RT_METRICS.xy, 0.0, 0.0)) +#define SMAASample(tex, coord) tex2D(tex, coord) +#define SMAASamplePoint(tex, coord) tex2D(tex, coord) +#define SMAASampleOffset(tex, coord, offset) tex2D(tex, coord + offset * SMAA_RT_METRICS.xy) +#define SMAA_FLATTEN [flatten] +#define SMAA_BRANCH [branch] +#endif +#if defined(SMAA_HLSL_4) || defined(SMAA_HLSL_4_1) +SamplerState LinearSampler { Filter = MIN_MAG_LINEAR_MIP_POINT; AddressU = Clamp; AddressV = Clamp; }; +SamplerState PointSampler { Filter = MIN_MAG_MIP_POINT; AddressU = Clamp; AddressV = Clamp; }; +#define SMAATexture2D(tex) Texture2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) tex.SampleLevel(LinearSampler, coord, 0) +#define SMAASampleLevelZeroPoint(tex, coord) tex.SampleLevel(PointSampler, coord, 0) +#define SMAASampleLevelZeroOffset(tex, coord, offset) tex.SampleLevel(LinearSampler, coord, 0, offset) +#define SMAASample(tex, coord) tex.Sample(LinearSampler, coord) +#define SMAASamplePoint(tex, coord) tex.Sample(PointSampler, coord) +#define SMAASampleOffset(tex, coord, offset) tex.Sample(LinearSampler, coord, offset) +#define SMAA_FLATTEN [flatten] +#define SMAA_BRANCH [branch] +#define SMAATexture2DMS2(tex) Texture2DMS tex +#define SMAALoad(tex, pos, sample) tex.Load(pos, sample) +#if defined(SMAA_HLSL_4_1) +#define SMAAGather(tex, coord) tex.Gather(LinearSampler, coord, 0) +#endif +#endif +#if defined(SMAA_GLSL_3) || defined(SMAA_GLSL_4) +#define SMAATexture2D(tex) sampler2D tex +#define SMAATexturePass2D(tex) tex +#define SMAASampleLevelZero(tex, coord) textureLod(tex, coord, 0.0) +#define SMAASampleLevelZeroPoint(tex, coord) textureLod(tex, coord, 0.0) +#define SMAASampleLevelZeroOffset(tex, coord, offset) textureLodOffset(tex, coord, 0.0, offset) +#define SMAASample(tex, coord) texture(tex, coord) +#define SMAASamplePoint(tex, coord) texture(tex, coord) +#define SMAASampleOffset(tex, coord, offset) texture(tex, coord, offset) +#define SMAA_FLATTEN +#define SMAA_BRANCH +#define lerp(a, b, t) mix(a, b, t) +#define saturate(a) clamp(a, 0.0, 1.0) +#if defined(SMAA_GLSL_4) +#define mad(a, b, c) fma(a, b, c) +#define SMAAGather(tex, coord) textureGather(tex, coord) +#else +#define mad(a, b, c) (a * b + c) +#endif +#define float2 vec2 +#define float3 vec3 +#define float4 vec4 +#define int2 ivec2 +#define int3 ivec3 +#define int4 ivec4 +#define bool2 bvec2 +#define bool3 bvec3 +#define bool4 bvec4 +#endif + +#if !defined(SMAA_HLSL_3) && !defined(SMAA_HLSL_4) && !defined(SMAA_HLSL_4_1) && !defined(SMAA_GLSL_3) && !defined(SMAA_GLSL_4) && !defined(SMAA_CUSTOM_SL) +#error you must define the shading language: SMAA_HLSL_*, SMAA_GLSL_* or SMAA_CUSTOM_SL +#endif + +//----------------------------------------------------------------------------- +// Misc functions + +/** + * Gathers current pixel, and the top-left neighbors. + */ +float3 SMAAGatherNeighbours(float2 texcoord, + float4 offset[3], + SMAATexture2D(tex)) { + #ifdef SMAAGather + return SMAAGather(tex, texcoord + SMAA_RT_METRICS.xy * float2(-0.5, -0.5)).grb; + #else + float P = SMAASamplePoint(tex, texcoord).r; + float Pleft = SMAASamplePoint(tex, offset[0].xy).r; + float Ptop = SMAASamplePoint(tex, offset[0].zw).r; + return float3(P, Pleft, Ptop); + #endif +} + +/** + * Adjusts the threshold by means of predication. + */ +float2 SMAACalculatePredicatedThreshold(float2 texcoord, + float4 offset[3], + SMAATexture2D(predicationTex)) { + float3 neighbours = SMAAGatherNeighbours(texcoord, offset, SMAATexturePass2D(predicationTex)); + float2 delta = abs(neighbours.xx - neighbours.yz); + float2 edges = step(SMAA_PREDICATION_THRESHOLD, delta); + return SMAA_PREDICATION_SCALE * SMAA_THRESHOLD * (1.0 - SMAA_PREDICATION_STRENGTH * edges); +} + +/** + * Conditional move: + */ +void SMAAMovc(bool2 cond, inout float2 variable, float2 value) { + SMAA_FLATTEN if (cond.x) variable.x = value.x; + SMAA_FLATTEN if (cond.y) variable.y = value.y; +} + +void SMAAMovc(bool4 cond, inout float4 variable, float4 value) { + SMAAMovc(cond.xy, variable.xy, value.xy); + SMAAMovc(cond.zw, variable.zw, value.zw); +} + + +#if SMAA_INCLUDE_VS +//----------------------------------------------------------------------------- +// Vertex Shaders + +/** + * Edge Detection Vertex Shader + */ +void SMAAEdgeDetectionVS(float2 texcoord, + out float4 offset[3]) { + offset[0] = mad(SMAA_RT_METRICS.xyxy, float4(-1.0, 0.0, 0.0, -1.0), texcoord.xyxy); + offset[1] = mad(SMAA_RT_METRICS.xyxy, float4( 1.0, 0.0, 0.0, 1.0), texcoord.xyxy); + offset[2] = mad(SMAA_RT_METRICS.xyxy, float4(-2.0, 0.0, 0.0, -2.0), texcoord.xyxy); +} + +/** + * Blend Weight Calculation Vertex Shader + */ +void SMAABlendingWeightCalculationVS(float2 texcoord, + out float2 pixcoord, + out float4 offset[3]) { + pixcoord = texcoord * SMAA_RT_METRICS.zw; + + // We will use these offsets for the searches later on (see @PSEUDO_GATHER4): + offset[0] = mad(SMAA_RT_METRICS.xyxy, float4(-0.25, -0.125, 1.25, -0.125), texcoord.xyxy); + offset[1] = mad(SMAA_RT_METRICS.xyxy, float4(-0.125, -0.25, -0.125, 1.25), texcoord.xyxy); + + // And these for the searches, they indicate the ends of the loops: + offset[2] = mad(SMAA_RT_METRICS.xxyy, + float4(-2.0, 2.0, -2.0, 2.0) * float(SMAA_MAX_SEARCH_STEPS), + float4(offset[0].xz, offset[1].yw)); +} + +/** + * Neighborhood Blending Vertex Shader + */ +void SMAANeighborhoodBlendingVS(float2 texcoord, + out float4 offset) { + offset = mad(SMAA_RT_METRICS.xyxy, float4( 1.0, 0.0, 0.0, 1.0), texcoord.xyxy); +} +#endif // SMAA_INCLUDE_VS + +#if SMAA_INCLUDE_PS +//----------------------------------------------------------------------------- +// Edge Detection Pixel Shaders (First Pass) + +/** + * Luma Edge Detection + * + * IMPORTANT NOTICE: luma edge detection requires gamma-corrected colors, and + * thus 'colorTex' should be a non-sRGB texture. + */ +float2 SMAALumaEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(colorTex) + #if SMAA_PREDICATION + , SMAATexture2D(predicationTex) + #endif + ) { + // Calculate the threshold: + #if SMAA_PREDICATION + float2 threshold = SMAACalculatePredicatedThreshold(texcoord, offset, SMAATexturePass2D(predicationTex)); + #else + float2 threshold = float2(SMAA_THRESHOLD, SMAA_THRESHOLD); + #endif + + // Calculate lumas: + float3 weights = float3(0.2126, 0.7152, 0.0722); + float L = dot(SMAASamplePoint(colorTex, texcoord).rgb, weights); + + float Lleft = dot(SMAASamplePoint(colorTex, offset[0].xy).rgb, weights); + float Ltop = dot(SMAASamplePoint(colorTex, offset[0].zw).rgb, weights); + + // We do the usual threshold: + float4 delta; + delta.xy = abs(L - float2(Lleft, Ltop)); + float2 edges = step(threshold, delta.xy); + + // Then discard if there is no edge: + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + // Calculate right and bottom deltas: + float Lright = dot(SMAASamplePoint(colorTex, offset[1].xy).rgb, weights); + float Lbottom = dot(SMAASamplePoint(colorTex, offset[1].zw).rgb, weights); + delta.zw = abs(L - float2(Lright, Lbottom)); + + // Calculate the maximum delta in the direct neighborhood: + float2 maxDelta = max(delta.xy, delta.zw); + + // Calculate left-left and top-top deltas: + float Lleftleft = dot(SMAASamplePoint(colorTex, offset[2].xy).rgb, weights); + float Ltoptop = dot(SMAASamplePoint(colorTex, offset[2].zw).rgb, weights); + delta.zw = abs(float2(Lleft, Ltop) - float2(Lleftleft, Ltoptop)); + + // Calculate the final maximum delta: + maxDelta = max(maxDelta.xy, delta.zw); + float finalDelta = max(maxDelta.x, maxDelta.y); + + // Local contrast adaptation: + edges.xy *= step(finalDelta, SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR * delta.xy); + + return edges; +} + +/** + * Color Edge Detection + * + * IMPORTANT NOTICE: color edge detection requires gamma-corrected colors, and + * thus 'colorTex' should be a non-sRGB texture. + */ +float2 SMAAColorEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(colorTex) + #if SMAA_PREDICATION + , SMAATexture2D(predicationTex) + #endif + ) { + // Calculate the threshold: + #if SMAA_PREDICATION + float2 threshold = SMAACalculatePredicatedThreshold(texcoord, offset, predicationTex); + #else + float2 threshold = float2(SMAA_THRESHOLD, SMAA_THRESHOLD); + #endif + + // Calculate color deltas: + float4 delta; + float3 C = SMAASamplePoint(colorTex, texcoord).rgb; + + float3 Cleft = SMAASamplePoint(colorTex, offset[0].xy).rgb; + float3 t = abs(C - Cleft); + delta.x = max(max(t.r, t.g), t.b); + + float3 Ctop = SMAASamplePoint(colorTex, offset[0].zw).rgb; + t = abs(C - Ctop); + delta.y = max(max(t.r, t.g), t.b); + + // We do the usual threshold: + float2 edges = step(threshold, delta.xy); + + // Then discard if there is no edge: + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + // Calculate right and bottom deltas: + float3 Cright = SMAASamplePoint(colorTex, offset[1].xy).rgb; + t = abs(C - Cright); + delta.z = max(max(t.r, t.g), t.b); + + float3 Cbottom = SMAASamplePoint(colorTex, offset[1].zw).rgb; + t = abs(C - Cbottom); + delta.w = max(max(t.r, t.g), t.b); + + // Calculate the maximum delta in the direct neighborhood: + float2 maxDelta = max(delta.xy, delta.zw); + + // Calculate left-left and top-top deltas: + float3 Cleftleft = SMAASamplePoint(colorTex, offset[2].xy).rgb; + t = abs(C - Cleftleft); + delta.z = max(max(t.r, t.g), t.b); + + float3 Ctoptop = SMAASamplePoint(colorTex, offset[2].zw).rgb; + t = abs(C - Ctoptop); + delta.w = max(max(t.r, t.g), t.b); + + // Calculate the final maximum delta: + maxDelta = max(maxDelta.xy, delta.zw); + float finalDelta = max(maxDelta.x, maxDelta.y); + + // Local contrast adaptation: + edges.xy *= step(finalDelta, SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR * delta.xy); + + return edges; +} + +/** + * Depth Edge Detection + */ +float2 SMAADepthEdgeDetectionPS(float2 texcoord, + float4 offset[3], + SMAATexture2D(depthTex)) { + float3 neighbours = SMAAGatherNeighbours(texcoord, offset, SMAATexturePass2D(depthTex)); + float2 delta = abs(neighbours.xx - float2(neighbours.y, neighbours.z)); + float2 edges = step(SMAA_DEPTH_THRESHOLD, delta); + + if (dot(edges, float2(1.0, 1.0)) == 0.0) + return float2(-2.0, -2.0); + + return edges; +} + +//----------------------------------------------------------------------------- +// Diagonal Search Functions + +#if !defined(SMAA_DISABLE_DIAG_DETECTION) + +/** + * Allows to decode two binary values from a bilinear-filtered access. + */ +float2 SMAADecodeDiagBilinearAccess(float2 e) { + // Bilinear access for fetching 'e' have a 0.25 offset, and we are + // interested in the R and G edges: + // + // +---G---+-------+ + // | x o R x | + // +-------+-------+ + // + // Then, if one of these edge is enabled: + // Red: (0.75 * X + 0.25 * 1) => 0.25 or 1.0 + // Green: (0.75 * 1 + 0.25 * X) => 0.75 or 1.0 + // + // This function will unpack the values (mad + mul + round): + // wolframalpha.com: round(x * abs(5 * x - 5 * 0.75)) plot 0 to 1 + e.r = e.r * abs(5.0 * e.r - 5.0 * 0.75); + return round(e); +} + +float4 SMAADecodeDiagBilinearAccess(float4 e) { + e.rb = e.rb * abs(5.0 * e.rb - 5.0 * 0.75); + return round(e); +} + +/** + * These functions allows to perform diagonal pattern searches. + */ +float2 SMAASearchDiag1(SMAATexture2D(edgesTex), float2 texcoord, float2 dir, out float2 e) { + float4 coord = float4(texcoord, -1.0, 1.0); + float3 t = float3(SMAA_RT_METRICS.xy, 1.0); + while (coord.z < float(SMAA_MAX_SEARCH_STEPS_DIAG - 1) && + coord.w > 0.9) { + coord.xyz = mad(t, float3(dir, 1.0), coord.xyz); + e = SMAASampleLevelZero(edgesTex, coord.xy).rg; + coord.w = dot(e, float2(0.5, 0.5)); + } + return coord.zw; +} + +float2 SMAASearchDiag2(SMAATexture2D(edgesTex), float2 texcoord, float2 dir, out float2 e) { + float4 coord = float4(texcoord, -1.0, 1.0); + coord.x += 0.25 * SMAA_RT_METRICS.x; // See @SearchDiag2Optimization + float3 t = float3(SMAA_RT_METRICS.xy, 1.0); + while (coord.z < float(SMAA_MAX_SEARCH_STEPS_DIAG - 1) && + coord.w > 0.9) { + coord.xyz = mad(t, float3(dir, 1.0), coord.xyz); + + // @SearchDiag2Optimization + // Fetch both edges at once using bilinear filtering: + e = SMAASampleLevelZero(edgesTex, coord.xy).rg; + e = SMAADecodeDiagBilinearAccess(e); + + // Non-optimized version: + // e.g = SMAASampleLevelZero(edgesTex, coord.xy).g; + // e.r = SMAASampleLevelZeroOffset(edgesTex, coord.xy, int2(1, 0)).r; + + coord.w = dot(e, float2(0.5, 0.5)); + } + return coord.zw; +} + +/** + * Similar to SMAAArea, this calculates the area corresponding to a certain + * diagonal distance and crossing edges 'e'. + */ +float2 SMAAAreaDiag(SMAATexture2D(areaTex), float2 dist, float2 e, float offset) { + float2 texcoord = mad(float2(SMAA_AREATEX_MAX_DISTANCE_DIAG, SMAA_AREATEX_MAX_DISTANCE_DIAG), e, dist); + + // We do a scale and bias for mapping to texel space: + texcoord = mad(SMAA_AREATEX_PIXEL_SIZE, texcoord, 0.5 * SMAA_AREATEX_PIXEL_SIZE); + + // Diagonal areas are on the second half of the texture: + texcoord.x += 0.5; + + // Move to proper place, according to the subpixel offset: + texcoord.y += SMAA_AREATEX_SUBTEX_SIZE * offset; + + // Do it! + return SMAA_AREATEX_SELECT(SMAASampleLevelZero(areaTex, texcoord)); +} + +/** + * This searches for diagonal patterns and returns the corresponding weights. + */ +float2 SMAACalculateDiagWeights(SMAATexture2D(edgesTex), SMAATexture2D(areaTex), float2 texcoord, float2 e, float4 subsampleIndices) { + float2 weights = float2(0.0, 0.0); + + // Search for the line ends: + float4 d; + float2 end; + if (e.r > 0.0) { + d.xz = SMAASearchDiag1(SMAATexturePass2D(edgesTex), texcoord, float2(-1.0, 1.0), end); + d.x += float(end.y > 0.9); + } else + d.xz = float2(0.0, 0.0); + d.yw = SMAASearchDiag1(SMAATexturePass2D(edgesTex), texcoord, float2(1.0, -1.0), end); + + SMAA_BRANCH + if (d.x + d.y > 2.0) { // d.x + d.y + 1 > 3 + // Fetch the crossing edges: + float4 coords = mad(float4(-d.x + 0.25, d.x, d.y, -d.y - 0.25), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + float4 c; + c.xy = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).rg; + c.zw = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).rg; + c.yxwz = SMAADecodeDiagBilinearAccess(c.xyzw); + + // Non-optimized version: + // float4 coords = mad(float4(-d.x, d.x, d.y, -d.y), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + // float4 c; + // c.x = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).g; + // c.y = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2( 0, 0)).r; + // c.z = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).g; + // c.w = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, -1)).r; + + // Merge crossing edges at each side into a single value: + float2 cc = mad(float2(2.0, 2.0), c.xz, c.yw); + + // Remove the crossing edge if we didn't found the end of the line: + SMAAMovc(bool2(step(0.9, d.zw)), cc, float2(0.0, 0.0)); + + // Fetch the areas for this line: + weights += SMAAAreaDiag(SMAATexturePass2D(areaTex), d.xy, cc, subsampleIndices.z); + } + + // Search for the line ends: + d.xz = SMAASearchDiag2(SMAATexturePass2D(edgesTex), texcoord, float2(-1.0, -1.0), end); + if (SMAASampleLevelZeroOffset(edgesTex, texcoord, int2(1, 0)).r > 0.0) { + d.yw = SMAASearchDiag2(SMAATexturePass2D(edgesTex), texcoord, float2(1.0, 1.0), end); + d.y += float(end.y > 0.9); + } else + d.yw = float2(0.0, 0.0); + + SMAA_BRANCH + if (d.x + d.y > 2.0) { // d.x + d.y + 1 > 3 + // Fetch the crossing edges: + float4 coords = mad(float4(-d.x, -d.x, d.y, d.y), SMAA_RT_METRICS.xyxy, texcoord.xyxy); + float4 c; + c.x = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1, 0)).g; + c.y = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2( 0, -1)).r; + c.zw = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, 0)).gr; + float2 cc = mad(float2(2.0, 2.0), c.xz, c.yw); + + // Remove the crossing edge if we didn't found the end of the line: + SMAAMovc(bool2(step(0.9, d.zw)), cc, float2(0.0, 0.0)); + + // Fetch the areas for this line: + weights += SMAAAreaDiag(SMAATexturePass2D(areaTex), d.xy, cc, subsampleIndices.w).gr; + } + + return weights; +} +#endif + +//----------------------------------------------------------------------------- +// Horizontal/Vertical Search Functions + +/** + * This allows to determine how much length should we add in the last step + * of the searches. It takes the bilinearly interpolated edge (see + * @PSEUDO_GATHER4), and adds 0, 1 or 2, depending on which edges and + * crossing edges are active. + */ +float SMAASearchLength(SMAATexture2D(searchTex), float2 e, float offset) { + // The texture is flipped vertically, with left and right cases taking half + // of the space horizontally: + float2 scale = SMAA_SEARCHTEX_SIZE * float2(0.5, -1.0); + float2 bias = SMAA_SEARCHTEX_SIZE * float2(offset, 1.0); + + // Scale and bias to access texel centers: + scale += float2(-1.0, 1.0); + bias += float2( 0.5, -0.5); + + // Convert from pixel coordinates to texcoords: + // (We use SMAA_SEARCHTEX_PACKED_SIZE because the texture is cropped) + scale *= 1.0 / SMAA_SEARCHTEX_PACKED_SIZE; + bias *= 1.0 / SMAA_SEARCHTEX_PACKED_SIZE; + + // Lookup the search texture: + return SMAA_SEARCHTEX_SELECT(SMAASampleLevelZero(searchTex, mad(scale, e, bias))); +} + +/** + * Horizontal/vertical search functions for the 2nd pass. + */ +float SMAASearchXLeft(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + /** + * @PSEUDO_GATHER4 + * This texcoord has been offset by (-0.25, -0.125) in the vertex shader to + * sample between edge, thus fetching four edges in a row. + * Sampling with different offsets in each direction allows to disambiguate + * which edges are active from the four fetched ones. + */ + float2 e = float2(0.0, 1.0); + while (texcoord.x > end && + e.g > 0.8281 && // Is there some edge not activated? + e.r == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(-float2(2.0, 0.0), SMAA_RT_METRICS.xy, texcoord); + } + + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.0), 3.25); + return mad(SMAA_RT_METRICS.x, offset, texcoord.x); + + // Non-optimized version: + // We correct the previous (-0.25, -0.125) offset we applied: + // texcoord.x += 0.25 * SMAA_RT_METRICS.x; + + // The searches are bias by 1, so adjust the coords accordingly: + // texcoord.x += SMAA_RT_METRICS.x; + + // Disambiguate the length added by the last step: + // texcoord.x += 2.0 * SMAA_RT_METRICS.x; // Undo last step + // texcoord.x -= SMAA_RT_METRICS.x * (255.0 / 127.0) * SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.0); + // return mad(SMAA_RT_METRICS.x, offset, texcoord.x); +} + +float SMAASearchXRight(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(0.0, 1.0); + while (texcoord.x < end && + e.g > 0.8281 && // Is there some edge not activated? + e.r == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(float2(2.0, 0.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.5), 3.25); + return mad(-SMAA_RT_METRICS.x, offset, texcoord.x); +} + +float SMAASearchYUp(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(1.0, 0.0); + while (texcoord.y > end && + e.r > 0.8281 && // Is there some edge not activated? + e.g == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(-float2(0.0, 2.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e.gr, 0.0), 3.25); + return mad(SMAA_RT_METRICS.y, offset, texcoord.y); +} + +float SMAASearchYDown(SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end) { + float2 e = float2(1.0, 0.0); + while (texcoord.y < end && + e.r > 0.8281 && // Is there some edge not activated? + e.g == 0.0) { // Or is there a crossing edge that breaks the line? + e = SMAASampleLevelZero(edgesTex, texcoord).rg; + texcoord = mad(float2(0.0, 2.0), SMAA_RT_METRICS.xy, texcoord); + } + float offset = mad(-(255.0 / 127.0), SMAASearchLength(SMAATexturePass2D(searchTex), e.gr, 0.5), 3.25); + return mad(-SMAA_RT_METRICS.y, offset, texcoord.y); +} + +/** + * Ok, we have the distance and both crossing edges. So, what are the areas + * at each side of current edge? + */ +float2 SMAAArea(SMAATexture2D(areaTex), float2 dist, float e1, float e2, float offset) { + // Rounding prevents precision errors of bilinear filtering: + float2 texcoord = mad(float2(SMAA_AREATEX_MAX_DISTANCE, SMAA_AREATEX_MAX_DISTANCE), round(4.0 * float2(e1, e2)), dist); + + // We do a scale and bias for mapping to texel space: + texcoord = mad(SMAA_AREATEX_PIXEL_SIZE, texcoord, 0.5 * SMAA_AREATEX_PIXEL_SIZE); + + // Move to proper place, according to the subpixel offset: + texcoord.y = mad(SMAA_AREATEX_SUBTEX_SIZE, offset, texcoord.y); + + // Do it! + return SMAA_AREATEX_SELECT(SMAASampleLevelZero(areaTex, texcoord)); +} + +//----------------------------------------------------------------------------- +// Corner Detection Functions + +void SMAADetectHorizontalCornerPattern(SMAATexture2D(edgesTex), inout float2 weights, float4 texcoord, float2 d) { + #if !defined(SMAA_DISABLE_CORNER_DETECTION) + float2 leftRight = step(d.xy, d.yx); + float2 rounding = (1.0 - SMAA_CORNER_ROUNDING_NORM) * leftRight; + + rounding /= leftRight.x + leftRight.y; // Reduce blending for pixels in the center of a line. + + float2 factor = float2(1.0, 1.0); + factor.x -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(0, 1)).r; + factor.x -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(1, 1)).r; + factor.y -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(0, -2)).r; + factor.y -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(1, -2)).r; + + weights *= saturate(factor); + #endif +} + +void SMAADetectVerticalCornerPattern(SMAATexture2D(edgesTex), inout float2 weights, float4 texcoord, float2 d) { + #if !defined(SMAA_DISABLE_CORNER_DETECTION) + float2 leftRight = step(d.xy, d.yx); + float2 rounding = (1.0 - SMAA_CORNER_ROUNDING_NORM) * leftRight; + + rounding /= leftRight.x + leftRight.y; + + float2 factor = float2(1.0, 1.0); + factor.x -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2( 1, 0)).g; + factor.x -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2( 1, 1)).g; + factor.y -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy, int2(-2, 0)).g; + factor.y -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw, int2(-2, 1)).g; + + weights *= saturate(factor); + #endif +} + +//----------------------------------------------------------------------------- +// Blending Weight Calculation Pixel Shader (Second Pass) + +float4 SMAABlendingWeightCalculationPS(float2 texcoord, + float2 pixcoord, + float4 offset[3], + SMAATexture2D(edgesTex), + SMAATexture2D(areaTex), + SMAATexture2D(searchTex), + float4 subsampleIndices) { // Just pass zero for SMAA 1x, see @SUBSAMPLE_INDICES. + float4 weights = float4(0.0, 0.0, 0.0, 0.0); + + float2 e = SMAASample(edgesTex, texcoord).rg; + + SMAA_BRANCH + if (e.g > 0.0) { // Edge at north + #if !defined(SMAA_DISABLE_DIAG_DETECTION) + // Diagonals have both north and west edges, so searching for them in + // one of the boundaries is enough. + weights.rg = SMAACalculateDiagWeights(SMAATexturePass2D(edgesTex), SMAATexturePass2D(areaTex), texcoord, e, subsampleIndices); + + // We give priority to diagonals, so if we find a diagonal we skip + // horizontal/vertical processing. + SMAA_BRANCH + if (weights.r == -weights.g) { // weights.r + weights.g == 0.0 + #endif + + float2 d; + + // Find the distance to the left: + float3 coords; + coords.x = SMAASearchXLeft(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[0].xy, offset[2].x); + coords.y = offset[1].y; // offset[1].y = texcoord.y - 0.25 * SMAA_RT_METRICS.y (@CROSSING_OFFSET) + d.x = coords.x; + + // Now fetch the left crossing edges, two at a time using bilinear + // filtering. Sampling at -0.25 (see @CROSSING_OFFSET) enables to + // discern what value each edge has: + float e1 = SMAASampleLevelZero(edgesTex, coords.xy).r; + + // Find the distance to the right: + coords.z = SMAASearchXRight(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[0].zw, offset[2].y); + d.y = coords.z; + + // We want the distances to be in pixel units (doing this here allow to + // better interleave arithmetic and memory accesses): + d = abs(round(mad(SMAA_RT_METRICS.zz, d, -pixcoord.xx))); + + // SMAAArea below needs a sqrt, as the areas texture is compressed + // quadratically: + float2 sqrt_d = sqrt(d); + + // Fetch the right crossing edges: + float e2 = SMAASampleLevelZeroOffset(edgesTex, coords.zy, int2(1, 0)).r; + + // Ok, we know how this pattern looks like, now it is time for getting + // the actual area: + weights.rg = SMAAArea(SMAATexturePass2D(areaTex), sqrt_d, e1, e2, subsampleIndices.y); + + // Fix corners: + coords.y = texcoord.y; + SMAADetectHorizontalCornerPattern(SMAATexturePass2D(edgesTex), weights.rg, coords.xyzy, d); + + #if !defined(SMAA_DISABLE_DIAG_DETECTION) + } else + e.r = 0.0; // Skip vertical processing. + #endif + } + + SMAA_BRANCH + if (e.r > 0.0) { // Edge at west + float2 d; + + // Find the distance to the top: + float3 coords; + coords.y = SMAASearchYUp(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[1].xy, offset[2].z); + coords.x = offset[0].x; // offset[1].x = texcoord.x - 0.25 * SMAA_RT_METRICS.x; + d.x = coords.y; + + // Fetch the top crossing edges: + float e1 = SMAASampleLevelZero(edgesTex, coords.xy).g; + + // Find the distance to the bottom: + coords.z = SMAASearchYDown(SMAATexturePass2D(edgesTex), SMAATexturePass2D(searchTex), offset[1].zw, offset[2].w); + d.y = coords.z; + + // We want the distances to be in pixel units: + d = abs(round(mad(SMAA_RT_METRICS.ww, d, -pixcoord.yy))); + + // SMAAArea below needs a sqrt, as the areas texture is compressed + // quadratically: + float2 sqrt_d = sqrt(d); + + // Fetch the bottom crossing edges: + float e2 = SMAASampleLevelZeroOffset(edgesTex, coords.xz, int2(0, 1)).g; + + // Get the area for this direction: + weights.ba = SMAAArea(SMAATexturePass2D(areaTex), sqrt_d, e1, e2, subsampleIndices.x); + + // Fix corners: + coords.x = texcoord.x; + SMAADetectVerticalCornerPattern(SMAATexturePass2D(edgesTex), weights.ba, coords.xyxz, d); + } + + return weights; +} + +//----------------------------------------------------------------------------- +// Neighborhood Blending Pixel Shader (Third Pass) + +float4 SMAANeighborhoodBlendingPS(float2 texcoord, + float4 offset, + SMAATexture2D(colorTex), + SMAATexture2D(blendTex) + #if SMAA_REPROJECTION + , SMAATexture2D(velocityTex) + #endif + ) { + // Fetch the blending weights for current pixel: + float4 a; + a.x = SMAASample(blendTex, offset.xy).a; // Right + a.y = SMAASample(blendTex, offset.zw).g; // Top + a.wz = SMAASample(blendTex, texcoord).xz; // Bottom / Left + + // Is there any blending weight with a value greater than 0.0? + SMAA_BRANCH + if (dot(a, float4(1.0, 1.0, 1.0, 1.0)) < 1e-5) { + float4 color = SMAASampleLevelZero(colorTex, texcoord); + + #if SMAA_REPROJECTION + float2 velocity = SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, texcoord)); + + // Pack velocity into the alpha channel: + color.a = sqrt(5.0 * length(velocity)); + #endif + + return color; + } else { + bool h = max(a.x, a.z) > max(a.y, a.w); // max(horizontal) > max(vertical) + + // Calculate the blending offsets: + float4 blendingOffset = float4(0.0, a.y, 0.0, a.w); + float2 blendingWeight = a.yw; + SMAAMovc(bool4(h, h, h, h), blendingOffset, float4(a.x, 0.0, a.z, 0.0)); + SMAAMovc(bool2(h, h), blendingWeight, a.xz); + blendingWeight /= dot(blendingWeight, float2(1.0, 1.0)); + + // Calculate the texture coordinates: + float4 blendingCoord = mad(blendingOffset, float4(SMAA_RT_METRICS.xy, -SMAA_RT_METRICS.xy), texcoord.xyxy); + + // We exploit bilinear filtering to mix current pixel with the chosen + // neighbor: + float4 color = blendingWeight.x * SMAASampleLevelZero(colorTex, blendingCoord.xy); + color += blendingWeight.y * SMAASampleLevelZero(colorTex, blendingCoord.zw); + + #if SMAA_REPROJECTION + // Antialias velocity for proper reprojection in a later stage: + float2 velocity = blendingWeight.x * SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, blendingCoord.xy)); + velocity += blendingWeight.y * SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, blendingCoord.zw)); + + // Pack velocity into the alpha channel: + color.a = sqrt(5.0 * length(velocity)); + #endif + + return color; + } +} + +//----------------------------------------------------------------------------- +// Temporal Resolve Pixel Shader (Optional Pass) + +float4 SMAAResolvePS(float2 texcoord, + SMAATexture2D(currentColorTex), + SMAATexture2D(previousColorTex) + #if SMAA_REPROJECTION + , SMAATexture2D(velocityTex) + #endif + ) { + #if SMAA_REPROJECTION + // Velocity is assumed to be calculated for motion blur, so we need to + // inverse it for reprojection: + float2 velocity = -SMAA_DECODE_VELOCITY(SMAASamplePoint(velocityTex, texcoord).rg); + + // Fetch current pixel: + float4 current = SMAASamplePoint(currentColorTex, texcoord); + + // Reproject current coordinates and fetch previous pixel: + float4 previous = SMAASamplePoint(previousColorTex, texcoord + velocity); + + // Attenuate the previous pixel if the velocity is different: + float delta = abs(current.a * current.a - previous.a * previous.a) / 5.0; + float weight = 0.5 * saturate(1.0 - sqrt(delta) * SMAA_REPROJECTION_WEIGHT_SCALE); + + // Blend the pixels according to the calculated weight: + return lerp(current, previous, weight); + #else + // Just blend the pixels: + float4 current = SMAASamplePoint(currentColorTex, texcoord); + float4 previous = SMAASamplePoint(previousColorTex, texcoord); + return lerp(current, previous, 0.5); + #endif +} + +//----------------------------------------------------------------------------- +// Separate Multisamples Pixel Shader (Optional Pass) + +#ifdef SMAALoad +void SMAASeparatePS(float4 position, + float2 texcoord, + out float4 target0, + out float4 target1, + SMAATexture2DMS2(colorTexMS)) { + int2 pos = int2(position.xy); + target0 = SMAALoad(colorTexMS, pos, 0); + target1 = SMAALoad(colorTexMS, pos, 1); +} +#endif + +//----------------------------------------------------------------------------- +#endif // SMAA_INCLUDE_PS + +layout(rgba8, binding = 0, set = 3) uniform image2D imgOutput; + +layout(binding = 1, set = 2) uniform sampler2D inputImg; +layout(binding = 3, set = 2) uniform sampler2D samplerBlend; +layout( binding = 2 ) uniform invResolution +{ + vec2 invResolution_data; +}; + +void main() { + vec2 loc = ivec2(gl_GlobalInvocationID.x * 4, gl_GlobalInvocationID.y * 4); + for(int i = 0; i < 4; i++) + { + for(int j = 0; j < 4; j++) + { + ivec2 texelCoord = ivec2(loc.x + i, loc.y + j); + vec2 coord = (texelCoord + vec2(0.5)) / invResolution_data; + vec2 pixCoord; + vec4 offset; + + SMAANeighborhoodBlendingVS(coord, offset); + + vec4 oColor = SMAANeighborhoodBlendingPS(coord, offset, inputImg, samplerBlend); + + imageStore(imgOutput, texelCoord, oColor); + } + } +} diff --git a/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaNeighbour.spv b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaNeighbour.spv new file mode 100644 index 00000000..fa0208f2 Binary files /dev/null and b/Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaNeighbour.spv differ diff --git a/Ryujinx.Graphics.Vulkan/Effects/SmaaConstants.cs b/Ryujinx.Graphics.Vulkan/Effects/SmaaConstants.cs new file mode 100644 index 00000000..a5f060f1 --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/SmaaConstants.cs @@ -0,0 +1,15 @@ +using System.Runtime.InteropServices; + +namespace Ryujinx.Graphics.Vulkan.Effects +{ + [StructLayout(LayoutKind.Sequential, Pack = 4)] + internal struct SmaaConstants + { + public int QualityLow; + public int QualityMedium; + public int QualityHigh; + public int QualityUltra; + public float Width; + public float Height; + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Effects/SmaaPostProcessingEffect.cs b/Ryujinx.Graphics.Vulkan/Effects/SmaaPostProcessingEffect.cs new file mode 100644 index 00000000..4dcdaa64 --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Effects/SmaaPostProcessingEffect.cs @@ -0,0 +1,314 @@ +using Ryujinx.Common; +using Ryujinx.Graphics.GAL; +using Ryujinx.Graphics.Shader; +using Ryujinx.Graphics.Shader.Translation; +using Silk.NET.Vulkan; +using System; +using Format = Ryujinx.Graphics.GAL.Format; + +namespace Ryujinx.Graphics.Vulkan.Effects +{ + internal partial class SmaaPostProcessingEffect : IPostProcessingEffect + { + public const int AreaWidth = 160; + public const int AreaHeight = 560; + public const int SearchWidth = 64; + public const int SearchHeight = 16; + + private readonly VulkanRenderer _renderer; + private ISampler _samplerLinear; + private SmaaConstants _specConstants; + private ShaderCollection _edgeProgram; + private ShaderCollection _blendProgram; + private ShaderCollection _neighbourProgram; + + private PipelineHelperShader _pipeline; + + private TextureView _outputTexture; + private TextureView _edgeOutputTexture; + private TextureView _blendOutputTexture; + private TextureView _areaTexture; + private TextureView _searchTexture; + private Device _device; + private bool _recreatePipelines; + private int _quality; + + public SmaaPostProcessingEffect(VulkanRenderer renderer, Device device, int quality) + { + _device = device; + _renderer = renderer; + _quality = quality; + + Initialize(); + } + + public int Quality + { + get => _quality; + set + { + _quality = value; + + _recreatePipelines = true; + } + } + + public void Dispose() + { + DeletePipelines(); + _samplerLinear?.Dispose(); + _outputTexture?.Dispose(); + _edgeOutputTexture?.Dispose(); + _blendOutputTexture?.Dispose(); + _areaTexture?.Dispose(); + _searchTexture?.Dispose(); + } + + private unsafe void RecreateShaders(int width, int height) + { + _recreatePipelines = false; + + DeletePipelines(); + _pipeline = new PipelineHelperShader(_renderer, _device); + + _pipeline.Initialize(); + + var edgeShader = EmbeddedResources.Read("Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaEdge.spv"); + var blendShader = EmbeddedResources.Read("Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaBlend.spv"); + var neighbourShader = EmbeddedResources.Read("Ryujinx.Graphics.Vulkan/Effects/Shaders/SmaaNeighbour.spv"); + + var edgeBindings = new ShaderBindings( + new[] { 2 }, + Array.Empty(), + new[] { 1 }, + new[] { 0 }); + + var blendBindings = new ShaderBindings( + new[] { 2 }, + Array.Empty(), + new[] { 1, 3, 4 }, + new[] { 0 }); + + var neighbourBindings = new ShaderBindings( + new[] { 2 }, + Array.Empty(), + new[] { 1, 3 }, + new[] { 0 }); + + _samplerLinear = _renderer.CreateSampler(GAL.SamplerCreateInfo.Create(MinFilter.Linear, MagFilter.Linear)); + + _specConstants = new SmaaConstants() + { + Width = width, + Height = height, + QualityLow = Quality == 0 ? 1 : 0, + QualityMedium = Quality == 1 ? 1 : 0, + QualityHigh = Quality == 2 ? 1 : 0, + QualityUltra = Quality == 3 ? 1 : 0, + }; + + var specInfo = new SpecDescription( + (0, SpecConstType.Int32), + (1, SpecConstType.Int32), + (2, SpecConstType.Int32), + (3, SpecConstType.Int32), + (4, SpecConstType.Float32), + (5, SpecConstType.Float32)); + + _edgeProgram = _renderer.CreateProgramWithMinimalLayout(new[] + { + new ShaderSource(edgeShader, edgeBindings, ShaderStage.Compute, TargetLanguage.Spirv) + }, new[] { specInfo }); + + _blendProgram = _renderer.CreateProgramWithMinimalLayout(new[] + { + new ShaderSource(blendShader, blendBindings, ShaderStage.Compute, TargetLanguage.Spirv) + }, new[] { specInfo }); + + _neighbourProgram = _renderer.CreateProgramWithMinimalLayout(new[] + { + new ShaderSource(neighbourShader, neighbourBindings, ShaderStage.Compute, TargetLanguage.Spirv) + }, new[] { specInfo }); + } + + public void DeletePipelines() + { + _pipeline?.Dispose(); + _edgeProgram?.Dispose(); + _blendProgram?.Dispose(); + _neighbourProgram?.Dispose(); + } + + private void Initialize() + { + var areaInfo = new TextureCreateInfo(AreaWidth, + AreaHeight, + 1, + 1, + 1, + 1, + 1, + 1, + Format.R8G8Unorm, + DepthStencilMode.Depth, + Target.Texture2D, + SwizzleComponent.Red, + SwizzleComponent.Green, + SwizzleComponent.Blue, + SwizzleComponent.Alpha); + + var searchInfo = new TextureCreateInfo(SearchWidth, + SearchHeight, + 1, + 1, + 1, + 1, + 1, + 1, + Format.R8Unorm, + DepthStencilMode.Depth, + Target.Texture2D, + SwizzleComponent.Red, + SwizzleComponent.Green, + SwizzleComponent.Blue, + SwizzleComponent.Alpha); + + var areaTexture = EmbeddedResources.Read("Ryujinx.Graphics.Vulkan/Effects/Textures/SmaaAreaTexture.bin"); + var searchTexture = EmbeddedResources.Read("Ryujinx.Graphics.Vulkan/Effects/Textures/SmaaSearchTexture.bin"); + + _areaTexture = _renderer.CreateTexture(areaInfo, 1) as TextureView; + _searchTexture = _renderer.CreateTexture(searchInfo, 1) as TextureView; + + _areaTexture.SetData(areaTexture); + _searchTexture.SetData(searchTexture); + } + + public TextureView Run(TextureView view, CommandBufferScoped cbs, int width, int height) + { + if (_recreatePipelines || _outputTexture == null || _outputTexture.Info.Width != view.Width || _outputTexture.Info.Height != view.Height) + { + RecreateShaders(view.Width, view.Height); + _outputTexture?.Dispose(); + _edgeOutputTexture?.Dispose(); + _blendOutputTexture?.Dispose(); + + var info = view.Info; + + if (view.Info.Format.IsBgr()) + { + info = new TextureCreateInfo(info.Width, + info.Height, + info.Depth, + info.Levels, + info.Samples, + info.BlockWidth, + info.BlockHeight, + info.BytesPerPixel, + info.Format, + info.DepthStencilMode, + info.Target, + info.SwizzleB, + info.SwizzleG, + info.SwizzleR, + info.SwizzleA); + } + + _outputTexture = _renderer.CreateTexture(info, view.ScaleFactor) as TextureView; + _edgeOutputTexture = _renderer.CreateTexture(info, view.ScaleFactor) as TextureView; + _blendOutputTexture = _renderer.CreateTexture(info, view.ScaleFactor) as TextureView; + } + + Span viewports = stackalloc GAL.Viewport[1]; + + viewports[0] = new GAL.Viewport( + new Rectangle(0, 0, view.Width, view.Height), + ViewportSwizzle.PositiveX, + ViewportSwizzle.PositiveY, + ViewportSwizzle.PositiveZ, + ViewportSwizzle.PositiveW, + 0f, + 1f); + + Span> scissors = stackalloc Rectangle[1]; + + scissors[0] = new Rectangle(0, 0, view.Width, view.Height); + + _renderer.HelperShader.Clear(_renderer, + _edgeOutputTexture.GetImageView(), + new float[] { 0, 0, 0, 1 }, + (uint)(ColorComponentFlags.RBit | ColorComponentFlags.GBit | ColorComponentFlags.BBit | ColorComponentFlags.ABit), + view.Width, + view.Height, + _edgeOutputTexture.VkFormat, + ComponentType.UnsignedInteger, + scissors[0]); + + _renderer.HelperShader.Clear(_renderer, + _blendOutputTexture.GetImageView(), + new float[] { 0, 0, 0, 1 }, + (uint)(ColorComponentFlags.RBit | ColorComponentFlags.GBit | ColorComponentFlags.BBit | ColorComponentFlags.ABit), + view.Width, + view.Height, + _blendOutputTexture.VkFormat, + ComponentType.UnsignedInteger, + scissors[0]); + + _renderer.Pipeline.TextureBarrier(); + + var dispatchX = BitUtils.DivRoundUp(view.Width, IPostProcessingEffect.LocalGroupSize); + var dispatchY = BitUtils.DivRoundUp(view.Height, IPostProcessingEffect.LocalGroupSize); + + // Edge pass + _pipeline.SetCommandBuffer(cbs); + _pipeline.SetProgram(_edgeProgram); + _pipeline.SetTextureAndSampler(ShaderStage.Compute, 1, view, _samplerLinear); + _pipeline.Specialize(_specConstants); + + ReadOnlySpan resolutionBuffer = stackalloc float[] { view.Width, view.Height }; + int rangeSize = resolutionBuffer.Length * sizeof(float); + var bufferHandle = _renderer.BufferManager.CreateWithHandle(_renderer, rangeSize, false); + + _renderer.BufferManager.SetData(bufferHandle, 0, resolutionBuffer); + var bufferRanges = new BufferRange(bufferHandle, 0, rangeSize); + _pipeline.SetUniformBuffers(stackalloc[] { new BufferAssignment(2, bufferRanges) }); + _pipeline.SetScissors(scissors); + _pipeline.SetViewports(viewports, false); + _pipeline.SetImage(0, _edgeOutputTexture, GAL.Format.R8G8B8A8Unorm); + _pipeline.DispatchCompute(dispatchX, dispatchY, 1); + _pipeline.ComputeBarrier(); + + // Blend pass + _pipeline.SetCommandBuffer(cbs); + _pipeline.SetProgram(_blendProgram); + _pipeline.Specialize(_specConstants); + _pipeline.SetTextureAndSampler(ShaderStage.Compute, 1, _edgeOutputTexture, _samplerLinear); + _pipeline.SetTextureAndSampler(ShaderStage.Compute, 3, _areaTexture, _samplerLinear); + _pipeline.SetTextureAndSampler(ShaderStage.Compute, 4, _searchTexture, _samplerLinear); + _pipeline.SetUniformBuffers(stackalloc[] { new BufferAssignment(2, bufferRanges) }); + _pipeline.SetScissors(scissors); + _pipeline.SetViewports(viewports, false); + _pipeline.SetImage(0, _blendOutputTexture, GAL.Format.R8G8B8A8Unorm); + _pipeline.DispatchCompute(dispatchX, dispatchY, 1); + _pipeline.ComputeBarrier(); + + // Neighbour pass + _pipeline.SetCommandBuffer(cbs); + _pipeline.SetProgram(_neighbourProgram); + _pipeline.Specialize(_specConstants); + _pipeline.SetTextureAndSampler(ShaderStage.Compute, 3, _blendOutputTexture, _samplerLinear); + _pipeline.SetTextureAndSampler(ShaderStage.Compute, 1, view, _samplerLinear); + _pipeline.SetUniformBuffers(stackalloc[] { new BufferAssignment(2, bufferRanges) }); + _pipeline.SetScissors(scissors); + _pipeline.SetViewports(viewports, false); + _pipeline.SetImage(0, _outputTexture, GAL.Format.R8G8B8A8Unorm); + _pipeline.DispatchCompute(dispatchX, dispatchY, 1); + _pipeline.ComputeBarrier(); + + _pipeline.Finish(); + + _renderer.BufferManager.Delete(bufferHandle); + + return _outputTexture; + } + } +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Effects/Textures/SmaaAreaTexture.bin b/Ryujinx.Graphics.Vulkan/Effects/Textures/SmaaAreaTexture.bin new file mode 100644 index 00000000..f4a7a1b4 Binary files /dev/null and b/Ryujinx.Graphics.Vulkan/Effects/Textures/SmaaAreaTexture.bin differ diff --git a/Ryujinx.Graphics.Vulkan/Effects/Textures/SmaaSearchTexture.bin b/Ryujinx.Graphics.Vulkan/Effects/Textures/SmaaSearchTexture.bin new file mode 100644 index 00000000..db5bf73f Binary files /dev/null and b/Ryujinx.Graphics.Vulkan/Effects/Textures/SmaaSearchTexture.bin differ diff --git a/Ryujinx.Graphics.Vulkan/NativeArray.cs b/Ryujinx.Graphics.Vulkan/NativeArray.cs index f7407439..3a851287 100644 --- a/Ryujinx.Graphics.Vulkan/NativeArray.cs +++ b/Ryujinx.Graphics.Vulkan/NativeArray.cs @@ -38,8 +38,11 @@ namespace Ryujinx.Graphics.Vulkan public void Dispose() { - Marshal.FreeHGlobal((IntPtr)Pointer); - Pointer = null; + if (Pointer != null) + { + Marshal.FreeHGlobal((IntPtr)Pointer); + Pointer = null; + } } } } diff --git a/Ryujinx.Graphics.Vulkan/PipelineBase.cs b/Ryujinx.Graphics.Vulkan/PipelineBase.cs index f779305d..583bb953 100644 --- a/Ryujinx.Graphics.Vulkan/PipelineBase.cs +++ b/Ryujinx.Graphics.Vulkan/PipelineBase.cs @@ -150,6 +150,28 @@ namespace Ryujinx.Graphics.Vulkan null); } + public void ComputeBarrier() + { + MemoryBarrier memoryBarrier = new MemoryBarrier() + { + SType = StructureType.MemoryBarrier, + SrcAccessMask = AccessFlags.MemoryReadBit | AccessFlags.MemoryWriteBit, + DstAccessMask = AccessFlags.MemoryReadBit | AccessFlags.MemoryWriteBit + }; + + Gd.Api.CmdPipelineBarrier( + CommandBuffer, + PipelineStageFlags.ComputeShaderBit, + PipelineStageFlags.AllCommandsBit, + 0, + 1, + new ReadOnlySpan(memoryBarrier), + 0, + ReadOnlySpan.Empty, + 0, + ReadOnlySpan.Empty); + } + public void BeginTransformFeedback(GAL.PrimitiveTopology topology) { _tfEnabled = true; @@ -803,6 +825,11 @@ namespace Ryujinx.Graphics.Vulkan _descriptorSetUpdater.SetImage(binding, image, imageFormat); } + public void SetImage(int binding, Auto image) + { + _descriptorSetUpdater.SetImage(binding, image); + } + public void SetIndexBuffer(BufferRange buffer, GAL.IndexType type) { if (buffer.Handle != BufferHandle.Null) diff --git a/Ryujinx.Graphics.Vulkan/Ryujinx.Graphics.Vulkan.csproj b/Ryujinx.Graphics.Vulkan/Ryujinx.Graphics.Vulkan.csproj index 87f14a6a..57e2240a 100644 --- a/Ryujinx.Graphics.Vulkan/Ryujinx.Graphics.Vulkan.csproj +++ b/Ryujinx.Graphics.Vulkan/Ryujinx.Graphics.Vulkan.csproj @@ -12,6 +12,17 @@ true + + + + + + + + + + + diff --git a/Ryujinx.Graphics.Vulkan/Window.cs b/Ryujinx.Graphics.Vulkan/Window.cs index a90a824d..5d6def3a 100644 --- a/Ryujinx.Graphics.Vulkan/Window.cs +++ b/Ryujinx.Graphics.Vulkan/Window.cs @@ -1,4 +1,5 @@ using Ryujinx.Graphics.GAL; +using Ryujinx.Graphics.Vulkan.Effects; using Silk.NET.Vulkan; using Silk.NET.Vulkan.Extensions.KHR; using System; @@ -29,6 +30,14 @@ namespace Ryujinx.Graphics.Vulkan private bool _vsyncEnabled; private bool _vsyncModeChanged; private VkFormat _format; + private AntiAliasing _currentAntiAliasing; + private bool _updateEffect; + private IPostProcessingEffect _effect; + private IScalingFilter _scalingFilter; + private bool _isLinear; + private float _scalingFilterLevel; + private bool _updateScalingFilter; + private ScalingFilter _currentScalingFilter; public unsafe Window(VulkanRenderer gd, SurfaceKHR surface, PhysicalDevice physicalDevice, Device device) { @@ -116,7 +125,7 @@ namespace Ryujinx.Graphics.Vulkan ImageFormat = surfaceFormat.Format, ImageColorSpace = surfaceFormat.ColorSpace, ImageExtent = extent, - ImageUsage = ImageUsageFlags.ColorAttachmentBit | ImageUsageFlags.TransferDstBit, + ImageUsage = ImageUsageFlags.ColorAttachmentBit | ImageUsageFlags.TransferDstBit | ImageUsageFlags.StorageBit, ImageSharingMode = SharingMode.Exclusive, ImageArrayLayers = 1, PreTransform = capabilities.CurrentTransform, @@ -280,6 +289,13 @@ namespace Ryujinx.Graphics.Vulkan var view = (TextureView)texture; + UpdateEffect(); + + if (_effect != null) + { + view = _effect.Run(view, cbs, _width, _height); + } + int srcX0, srcX1, srcY0, srcY1; float scale = view.ScaleFactor; @@ -315,6 +331,18 @@ namespace Ryujinx.Graphics.Vulkan if (ScreenCaptureRequested) { + if (_effect != null) + { + _gd.CommandBufferPool.Return( + cbs, + null, + stackalloc[] { PipelineStageFlags.ColorAttachmentOutputBit }, + null); + _gd.FlushAllCommands(); + cbs.GetFence().Wait(); + cbs = _gd.CommandBufferPool.Rent(); + } + CaptureFrame(view, srcX0, srcY0, srcX1 - srcX0, srcY1 - srcY0, view.Info.Format.IsBgr(), crop.FlipX, crop.FlipY); ScreenCaptureRequested = false; @@ -335,20 +363,36 @@ namespace Ryujinx.Graphics.Vulkan int dstY0 = crop.FlipY ? dstPaddingY : _height - dstPaddingY; int dstY1 = crop.FlipY ? _height - dstPaddingY : dstPaddingY; - _gd.HelperShader.BlitColor( - _gd, - cbs, - view, - _swapchainImageViews[nextImage], - _width, - _height, - 1, - _format, - false, - new Extents2D(srcX0, srcY0, srcX1, srcY1), - new Extents2D(dstX0, dstY1, dstX1, dstY0), - true, - true); + if (_scalingFilter != null) + { + _scalingFilter.Run( + view, + cbs, + _swapchainImageViews[nextImage], + _format, + _width, + _height, + new Extents2D(srcX0, srcY0, srcX1, srcY1), + new Extents2D(dstX0, dstY0, dstX1, dstY1) + ); + } + else + { + _gd.HelperShader.BlitColor( + _gd, + cbs, + view, + _swapchainImageViews[nextImage], + _width, + _height, + 1, + _format, + false, + new Extents2D(srcX0, srcY0, srcX1, srcY1), + new Extents2D(dstX0, dstY1, dstX1, dstY0), + _isLinear, + true); + } Transition( cbs.CommandBuffer, @@ -387,6 +431,95 @@ namespace Ryujinx.Graphics.Vulkan } } + public override void SetAntiAliasing(AntiAliasing effect) + { + if (_currentAntiAliasing == effect && _effect != null) + { + return; + } + + _currentAntiAliasing = effect; + + _updateEffect = true; + } + + public override void SetScalingFilter(ScalingFilter type) + { + if (_currentScalingFilter == type && _effect != null) + { + return; + } + + _currentScalingFilter = type; + + _updateScalingFilter = true; + } + + private void UpdateEffect() + { + if (_updateEffect) + { + _updateEffect = false; + + switch (_currentAntiAliasing) + { + case AntiAliasing.Fxaa: + _effect?.Dispose(); + _effect = new FxaaPostProcessingEffect(_gd, _device); + break; + case AntiAliasing.None: + _effect?.Dispose(); + _effect = null; + break; + case AntiAliasing.SmaaLow: + case AntiAliasing.SmaaMedium: + case AntiAliasing.SmaaHigh: + case AntiAliasing.SmaaUltra: + var quality = _currentAntiAliasing - AntiAliasing.SmaaLow; + if (_effect is SmaaPostProcessingEffect smaa) + { + smaa.Quality = quality; + } + else + { + _effect?.Dispose(); + _effect = new SmaaPostProcessingEffect(_gd, _device, quality); + } + break; + } + } + + if (_updateScalingFilter) + { + _updateScalingFilter = false; + + switch (_currentScalingFilter) + { + case ScalingFilter.Bilinear: + case ScalingFilter.Nearest: + _scalingFilter?.Dispose(); + _scalingFilter = null; + _isLinear = _currentScalingFilter == ScalingFilter.Bilinear; + break; + case ScalingFilter.Fsr: + if (_scalingFilter is not FsrScalingFilter) + { + _scalingFilter?.Dispose(); + _scalingFilter = new FsrScalingFilter(_gd, _device); + } + + _scalingFilter.Level = _scalingFilterLevel; + break; + } + } + } + + public override void SetScalingFilterLevel(float level) + { + _scalingFilterLevel = level; + _updateScalingFilter = true; + } + private unsafe void Transition( CommandBuffer commandBuffer, Image image, @@ -456,8 +589,10 @@ namespace Ryujinx.Graphics.Vulkan } _gd.SwapchainApi.DestroySwapchain(_device, _swapchain, null); - } + + _effect?.Dispose(); + _scalingFilter?.Dispose(); } } diff --git a/Ryujinx.Graphics.Vulkan/WindowBase.cs b/Ryujinx.Graphics.Vulkan/WindowBase.cs index 651fe7c1..0a365e8f 100644 --- a/Ryujinx.Graphics.Vulkan/WindowBase.cs +++ b/Ryujinx.Graphics.Vulkan/WindowBase.cs @@ -11,5 +11,8 @@ namespace Ryujinx.Graphics.Vulkan public abstract void Present(ITexture texture, ImageCrop crop, Action swapBuffersCallback); public abstract void SetSize(int width, int height); public abstract void ChangeVSyncMode(bool vsyncEnabled); + public abstract void SetAntiAliasing(AntiAliasing effect); + public abstract void SetScalingFilter(ScalingFilter scalerType); + public abstract void SetScalingFilterLevel(float scale); } } \ No newline at end of file diff --git a/Ryujinx.Ui.Common/Configuration/ConfigurationFileFormat.cs b/Ryujinx.Ui.Common/Configuration/ConfigurationFileFormat.cs index 226b5933..e9aec04b 100644 --- a/Ryujinx.Ui.Common/Configuration/ConfigurationFileFormat.cs +++ b/Ryujinx.Ui.Common/Configuration/ConfigurationFileFormat.cs @@ -14,7 +14,7 @@ namespace Ryujinx.Ui.Common.Configuration /// /// The current version of the file format /// - public const int CurrentVersion = 43; + public const int CurrentVersion = 44; /// /// Version of the configuration file format @@ -51,6 +51,21 @@ namespace Ryujinx.Ui.Common.Configuration /// public AspectRatio AspectRatio { get; set; } + /// + /// Applies anti-aliasing to the renderer. + /// + public AntiAliasing AntiAliasing { get; set; } + + /// + /// Sets the framebuffer upscaling type. + /// + public ScalingFilter ScalingFilter { get; set; } + + /// + /// Sets the framebuffer upscaling level. + /// + public int ScalingFilterLevel { get; set; } + /// /// Dumps shaders in this local directory /// diff --git a/Ryujinx.Ui.Common/Configuration/ConfigurationState.cs b/Ryujinx.Ui.Common/Configuration/ConfigurationState.cs index f193b157..bcdd2e70 100644 --- a/Ryujinx.Ui.Common/Configuration/ConfigurationState.cs +++ b/Ryujinx.Ui.Common/Configuration/ConfigurationState.cs @@ -433,6 +433,21 @@ namespace Ryujinx.Ui.Common.Configuration /// public ReactiveObject GraphicsBackend { get; private set; } + /// + /// Applies anti-aliasing to the renderer. + /// + public ReactiveObject AntiAliasing { get; private set; } + + /// + /// Sets the framebuffer upscaling type. + /// + public ReactiveObject ScalingFilter { get; private set; } + + /// + /// Sets the framebuffer upscaling level. + /// + public ReactiveObject ScalingFilterLevel { get; private set; } + /// /// Preferred GPU /// @@ -463,6 +478,12 @@ namespace Ryujinx.Ui.Common.Configuration PreferredGpu.Event += static (sender, e) => LogValueChange(sender, e, nameof(PreferredGpu)); EnableMacroHLE = new ReactiveObject(); EnableMacroHLE.Event += static (sender, e) => LogValueChange(sender, e, nameof(EnableMacroHLE)); + AntiAliasing = new ReactiveObject(); + AntiAliasing.Event += static (sender, e) => LogValueChange(sender, e, nameof(AntiAliasing)); + ScalingFilter = new ReactiveObject(); + ScalingFilter.Event += static (sender, e) => LogValueChange(sender, e, nameof(ScalingFilter)); + ScalingFilterLevel = new ReactiveObject(); + ScalingFilterLevel.Event += static (sender, e) => LogValueChange(sender, e, nameof(ScalingFilterLevel)); } } @@ -540,6 +561,9 @@ namespace Ryujinx.Ui.Common.Configuration ResScaleCustom = Graphics.ResScaleCustom, MaxAnisotropy = Graphics.MaxAnisotropy, AspectRatio = Graphics.AspectRatio, + AntiAliasing = Graphics.AntiAliasing, + ScalingFilter = Graphics.ScalingFilter, + ScalingFilterLevel = Graphics.ScalingFilterLevel, GraphicsShadersDumpPath = Graphics.ShadersDumpPath, LoggingEnableDebug = Logger.EnableDebug, LoggingEnableStub = Logger.EnableStub, @@ -651,6 +675,9 @@ namespace Ryujinx.Ui.Common.Configuration Graphics.EnableShaderCache.Value = true; Graphics.EnableTextureRecompression.Value = false; Graphics.EnableMacroHLE.Value = true; + Graphics.AntiAliasing.Value = AntiAliasing.None; + Graphics.ScalingFilter.Value = ScalingFilter.Bilinear; + Graphics.ScalingFilterLevel.Value = 80; System.EnablePtc.Value = true; System.EnableInternetAccess.Value = false; System.EnableFsIntegrityChecks.Value = true; @@ -1208,6 +1235,17 @@ namespace Ryujinx.Ui.Common.Configuration configurationFileFormat.UseHypervisor = true; } + if (configurationFileFormat.Version < 44) + { + Ryujinx.Common.Logging.Logger.Warning?.Print(LogClass.Application, $"Outdated configuration version {configurationFileFormat.Version}, migrating to version 42."); + + configurationFileFormat.AntiAliasing = AntiAliasing.None; + configurationFileFormat.ScalingFilter = ScalingFilter.Bilinear; + configurationFileFormat.ScalingFilterLevel = 80; + + configurationFileUpdated = true; + } + Logger.EnableFileLog.Value = configurationFileFormat.EnableFileLog; Graphics.ResScale.Value = configurationFileFormat.ResScale; Graphics.ResScaleCustom.Value = configurationFileFormat.ResScaleCustom; @@ -1217,6 +1255,9 @@ namespace Ryujinx.Ui.Common.Configuration Graphics.BackendThreading.Value = configurationFileFormat.BackendThreading; Graphics.GraphicsBackend.Value = configurationFileFormat.GraphicsBackend; Graphics.PreferredGpu.Value = configurationFileFormat.PreferredGpu; + Graphics.AntiAliasing.Value = configurationFileFormat.AntiAliasing; + Graphics.ScalingFilter.Value = configurationFileFormat.ScalingFilter; + Graphics.ScalingFilterLevel.Value = configurationFileFormat.ScalingFilterLevel; Logger.EnableDebug.Value = configurationFileFormat.LoggingEnableDebug; Logger.EnableStub.Value = configurationFileFormat.LoggingEnableStub; Logger.EnableInfo.Value = configurationFileFormat.LoggingEnableInfo; diff --git a/Ryujinx/Ui/RendererWidgetBase.cs b/Ryujinx/Ui/RendererWidgetBase.cs index 4bf2a70f..957bbcd5 100644 --- a/Ryujinx/Ui/RendererWidgetBase.cs +++ b/Ryujinx/Ui/RendererWidgetBase.cs @@ -27,6 +27,7 @@ namespace Ryujinx.Ui using Image = SixLabors.ImageSharp.Image; using Key = Input.Key; using Switch = HLE.Switch; + using ScalingFilter = Graphics.GAL.ScalingFilter; public abstract class RendererWidgetBase : DrawingArea { @@ -116,6 +117,21 @@ namespace Ryujinx.Ui _lastCursorMoveTime = Stopwatch.GetTimestamp(); ConfigurationState.Instance.HideCursorOnIdle.Event += HideCursorStateChanged; + ConfigurationState.Instance.Graphics.AntiAliasing.Event += UpdateAnriAliasing; + ConfigurationState.Instance.Graphics.ScalingFilter.Event += UpdateScalingFilter; + ConfigurationState.Instance.Graphics.ScalingFilterLevel.Event += UpdateScalingFilterLevel; + } + + private void UpdateScalingFilterLevel(object sender, ReactiveEventArgs e) + { + Renderer.Window.SetScalingFilter((ScalingFilter)ConfigurationState.Instance.Graphics.ScalingFilter.Value); + Renderer.Window.SetScalingFilterLevel(ConfigurationState.Instance.Graphics.ScalingFilterLevel.Value); + } + + private void UpdateScalingFilter(object sender, ReactiveEventArgs e) + { + Renderer.Window.SetScalingFilter((ScalingFilter)ConfigurationState.Instance.Graphics.ScalingFilter.Value); + Renderer.Window.SetScalingFilterLevel(ConfigurationState.Instance.Graphics.ScalingFilterLevel.Value); } public abstract void InitializeRenderer(); @@ -149,11 +165,19 @@ namespace Ryujinx.Ui private void Renderer_Destroyed(object sender, EventArgs e) { ConfigurationState.Instance.HideCursorOnIdle.Event -= HideCursorStateChanged; + ConfigurationState.Instance.Graphics.AntiAliasing.Event -= UpdateAnriAliasing; + ConfigurationState.Instance.Graphics.ScalingFilter.Event -= UpdateScalingFilter; + ConfigurationState.Instance.Graphics.ScalingFilterLevel.Event -= UpdateScalingFilterLevel; NpadManager.Dispose(); Dispose(); } + private void UpdateAnriAliasing(object sender, ReactiveEventArgs e) + { + Renderer?.Window.SetAntiAliasing((Graphics.GAL.AntiAliasing)e.NewValue); + } + protected override bool OnMotionNotifyEvent(EventMotion evnt) { if (_hideCursorOnIdle) @@ -394,6 +418,10 @@ namespace Ryujinx.Ui Device.Gpu.Renderer.Initialize(_glLogLevel); + Renderer.Window.SetAntiAliasing((Graphics.GAL.AntiAliasing)ConfigurationState.Instance.Graphics.AntiAliasing.Value); + Renderer.Window.SetScalingFilter((Graphics.GAL.ScalingFilter)ConfigurationState.Instance.Graphics.ScalingFilter.Value); + Renderer.Window.SetScalingFilterLevel(ConfigurationState.Instance.Graphics.ScalingFilterLevel.Value); + _gpuBackendName = GetGpuBackendName(); _gpuVendorName = GetGpuVendorName(); diff --git a/Ryujinx/Ui/Windows/SettingsWindow.cs b/Ryujinx/Ui/Windows/SettingsWindow.cs index 220bb82a..61af7d39 100644 --- a/Ryujinx/Ui/Windows/SettingsWindow.cs +++ b/Ryujinx/Ui/Windows/SettingsWindow.cs @@ -95,10 +95,14 @@ namespace Ryujinx.Ui.Windows [GUI] Entry _graphicsShadersDumpPath; [GUI] ComboBoxText _anisotropy; [GUI] ComboBoxText _aspectRatio; + [GUI] ComboBoxText _antiAliasing; + [GUI] ComboBoxText _scalingFilter; [GUI] ComboBoxText _graphicsBackend; [GUI] ComboBoxText _preferredGpu; [GUI] ComboBoxText _resScaleCombo; [GUI] Entry _resScaleText; + [GUI] Adjustment _scalingFilterLevel; + [GUI] Scale _scalingFilterSlider; [GUI] ToggleButton _configureController1; [GUI] ToggleButton _configureController2; [GUI] ToggleButton _configureController3; @@ -139,6 +143,7 @@ namespace Ryujinx.Ui.Windows _systemTimeZoneEntry.FocusOutEvent += TimeZoneEntry_FocusOut; _resScaleCombo.Changed += (sender, args) => _resScaleText.Visible = _resScaleCombo.ActiveId == "-1"; + _scalingFilter.Changed += (sender, args) => _scalingFilterSlider.Visible = _scalingFilter.ActiveId == "2"; _galThreading.Changed += (sender, args) => { if (_galThreading.ActiveId != ConfigurationState.Instance.Graphics.BackendThreading.Value.ToString()) @@ -338,6 +343,8 @@ namespace Ryujinx.Ui.Windows _anisotropy.SetActiveId(ConfigurationState.Instance.Graphics.MaxAnisotropy.Value.ToString()); _aspectRatio.SetActiveId(((int)ConfigurationState.Instance.Graphics.AspectRatio.Value).ToString()); _graphicsBackend.SetActiveId(((int)ConfigurationState.Instance.Graphics.GraphicsBackend.Value).ToString()); + _antiAliasing.SetActiveId(((int)ConfigurationState.Instance.Graphics.AntiAliasing.Value).ToString()); + _scalingFilter.SetActiveId(((int)ConfigurationState.Instance.Graphics.ScalingFilter.Value).ToString()); UpdatePreferredGpuComboBox(); @@ -345,7 +352,9 @@ namespace Ryujinx.Ui.Windows _custThemePath.Buffer.Text = ConfigurationState.Instance.Ui.CustomThemePath; _resScaleText.Buffer.Text = ConfigurationState.Instance.Graphics.ResScaleCustom.Value.ToString(); + _scalingFilterLevel.Value = ConfigurationState.Instance.Graphics.ScalingFilterLevel.Value; _resScaleText.Visible = _resScaleCombo.ActiveId == "-1"; + _scalingFilterSlider.Visible = _scalingFilter.ActiveId == "2"; _graphicsShadersDumpPath.Buffer.Text = ConfigurationState.Instance.Graphics.ShadersDumpPath; _fsLogSpinAdjustment.Value = ConfigurationState.Instance.System.FsGlobalAccessLogMode; _systemTimeOffset = ConfigurationState.Instance.System.SystemTimeOffset; @@ -605,6 +614,9 @@ namespace Ryujinx.Ui.Windows ConfigurationState.Instance.Graphics.ResScale.Value = int.Parse(_resScaleCombo.ActiveId); ConfigurationState.Instance.Graphics.ResScaleCustom.Value = resScaleCustom; ConfigurationState.Instance.System.AudioVolume.Value = (float)_audioVolumeSlider.Value / 100.0f; + ConfigurationState.Instance.Graphics.AntiAliasing.Value = Enum.Parse(_antiAliasing.ActiveId); + ConfigurationState.Instance.Graphics.ScalingFilter.Value = Enum.Parse(_scalingFilter.ActiveId); + ConfigurationState.Instance.Graphics.ScalingFilterLevel.Value = (int)_scalingFilterLevel.Value; _previousVolumeLevel = ConfigurationState.Instance.System.AudioVolume.Value; diff --git a/Ryujinx/Ui/Windows/SettingsWindow.glade b/Ryujinx/Ui/Windows/SettingsWindow.glade index e39be81a..c19c1db9 100644 --- a/Ryujinx/Ui/Windows/SettingsWindow.glade +++ b/Ryujinx/Ui/Windows/SettingsWindow.glade @@ -40,6 +40,13 @@ True True + + 0 + 101 + 1 + 5 + 1 + False Ryujinx - Settings @@ -2152,6 +2159,118 @@ 3 + + + True + False + 5 + 5 + + + True + False + Applies a final effect to the game render + Post Processing Effect: + + + False + True + 5 + 0 + + + + + True + False + Applies anti-aliasing to the game render + 1 + + None + FXAA + SMAA Low + SMAA Medium + SMAA High + SMAA Ultra + + + + False + True + 1 + + + + + False + True + 5 + 4 + + + + + 100 + True + False + 5 + 5 + + + True + False + Enables Framebuffer Upscaling + Upscale: + + + False + True + 5 + 0 + + + + + True + False + Enables Framebuffer Upscaling + 1 + + Bilinear + Nearest + FSR + + + + False + True + 1 + + + + + 200 + True + True + 5 + _scalingFilterLevel + 1 + right + + + False + True + 3 + + + + + False + True + 5 + 5 + + True @@ -2197,7 +2316,7 @@ False True 5 - 4 + 6 @@ -2246,7 +2365,7 @@ False True 5 - 5 + 7